Path: utzoo!utgpu!jarvis.csri.toronto.edu!mailrus!ames!apple!oliveb!pyramid!prls!philabs!phri!marob!cowan
From: cowan@marob.masa.com (John Cowan)
Newsgroups: comp.lang.c
Subject: switch statements (was: forward references in typedefs)
Message-ID: <24CDEE4B.4647@marob.masa.com>
Date: 26 Jul 89 17:00:57 GMT
References: <55480@tut.cis.ohio-state.edu> <1989Jul20.152935.14872@utzoo.uucp> <24CB9E07.9547@marob.masa.com> <18729@mimsy.UUCP>
Reply-To: cowan@marob.masa.com (John Cowan)
Organization: ESCC,  New York City
Lines: 36

In article <18729@mimsy.UUCP> chris@mimsy.UUCP (Chris Torek) writes:
>In article <24CB9E07.9547@marob.masa.com> cowan@marob.masa.com (John Cowan)
>writes:
>>... As someone (Chris Torek?) said earlier in this group, "all 'struct'
>>pointers must 'smell' the same."  If this behavior is not guaranteed,
>>forward references to structs would be unimplementable in one pass.
>
>It was Doug Gwyn, but this is otherwise correct.

Sorry about that.  I, like Usenet, tend to lose my memory every six months.
:-)

>>Ufcawss, if you talk to the people who wrote VMS C, they'll tell you that
>>all one-pass implementations of C are unacceptable!  (This has something to
>>do with generating good code for the 'switch' statement.)
>
>They are both right and wrong.  One-pass code generation means there
>are few opportunities for optimsiation; but generating good switches
>is easy.  Simply emit a branch at the top of the switch...

I looked up the complaint, and they don't like this unconditional branch.
It involves jumping around the whole of a potentially large switch
statement, and then back up to somewhere within the code.  In a paging
architecture with small pages (like VMS: 512 bytes/page), this may
involve an unnecessary page fault just to get to the jump table.

A two-pass compiler allows putting the jump table at the top.  The VMS
compiler actually emits most code on the first pass, watching in the mean
time for switch statements and a few other constructs, and then runs a
second pass to fix these constructs only if they were detected.  He who
eschews 'switch' wins fast compiles under VMS.
-- 
Internet/Smail: cowan@marob.masa.com	Dumb: uunet!hombre!marob!cowan
Fidonet:  JOHN COWAN of 1:107/711	Magpie: JOHN COWAN, (212) 420-0527
		Charles li reis, nostre emperesdre magnes
		Set anz toz pleins at estet in Espagne.