NetNews Usenet Archive 1992 #20/NN_1992_20.iso/spool/comp/sys/hp/10073

Internet Message Format | 1992-09-07 | 2.4 KB

Path: sparky!uunet!gatech!destroyer!caen!spool.mu.edu!olivea!sgigate!odin!fido!shankar
From: [e-mail] (Shankar Unni)
Newsgroups: comp.sys.hp
Subject: Re: that HPUX cc compiler
Message-ID: <[e-mail]>
Date: 4 Sep 92 21:44:06 GMT
References: <[e-mail]> <[e-mail]> <[e-mail]>
Sender: [e-mail] (Usenet News Admin)
Organization: Silicon Graphics, Inc.
Lines: 44

In article <[e-mail]> [e-mail] (Claude Lecommandeur) writes:

>    I have seen this behaviour before, i mistakenly gave the actual structure
>instead of its address as a parameter to a function call and it worked when
>compiled without -O and broke with -O. (it's hard to debug because it works
>with -g). In this case, it is definetely not the optimiser which is broken,
>but the compiler itself. It looks like it tries to be too much clever.

No, it's not - it's doing a perfectly reasonable thing.

You are passing a struct by value to a function whose prototype is not
declared (I'm talking K&R compiles here for the moment). Since that's
a perfectly legal thing to do, the compiler does so. It just happens
that the way the HP compiler does it is to pass in the address of the
structure to the callee, *which is supposed to make a local copy
immediately* (automatically, that is, if the parameter is declared
correctly as a struct by value).

If the target declares it as a struct by reference, you *could* have a
situation where the struct passed by value gets modified by the callee.
As is the case with the fread call.

However, back at the point of call, the optimizer is pretty confident
that the struct is not modified by the call (hey, it was passed by
value, wasn't it?). Which is why it thinks the assignment is
loop-invariant.

In this case, I agree that it may have been wiser to make the copy at
the point of call rather than at the point of entry - that way, the
above program would not have worked either optimized or unoptimized,
and the user would have been alerted.

Still, all these are just ordinary implementation decisions, and just
happen to have odd consequences if you break the rules in a way that
the compiler can't detect.

Answer(s):

  - (pre-ANSI): Lint it once, Lint it twice, Lint it once again....
  - (brave new ANSI world): cc -Aa, *and use those **** prototypes!!*

--
Shankar Unni                E-Mail:    [e-mail]
Silicon Graphics Inc.            Phone:    [phone]