Regarding calculation of free memory

It sounds like your 0xeeeee method you outline is perfectly suited for your purposes. Even if 0xeeeee is actually part of your real data that happens to be in the stack, consider that it's highly unlikely that

That's application specific. Look at your operating scenarios. Ask yourself if the pattern you've selected is likely to be real data or not.

See #1 reply.

Don't know but... This method is commonly used and usually works very well. A good white-box test designed to stress the stack usage will give you the real info you need to accurately size your RAM.

IMO, it's like gambling when you size how much stack margin you want to add. If you believe your SW tests are "NASA grade" (nearly exhaustive), you will not need to add much stack margin. If you don't do SW testing, it's a roll of the dice. You probably fit in the middle of the two extremes.

Good way is the way you outlined for run-time stack checking. Maybe some CPUs feature some sort of built-in high water mark stack pointer register. That would be nice but I don't know if it exists.

JJS

I believe the traditional value is "0xDEADBEEF" :)

Which has the same problem. But really, if you pick a pattern that's unlikely to turn up by coincidence, you'll come close enough to the right answer for all practical purposes. Don't use, say, 00000000, but beyond that it doesn't much matter.

There is also a slight advantage to using this value because it is odd. On many processors using an odd address for larger than byte size data will thow an exception. As a result uninitialised data on the stack used as pointers can be caught more often.

The bigest problem I have though is that I am a vegetarian so I prefer

Peter

Seen in various 16 bit codes: #define STACKPAT 0x55aa // Stack fill value for high water mark checking.

If you want your appliance to be totally reliable, avoid using a heap at all. Free memory calculation then becomes rather simple [1]. Logic: heaps are (usually) non-deterministic and *will* cause eventual malloc failures due to fragmentation [2].

Steve

Not so. In Fortran 77, it was possible to allocate all memory statically, but it isn't generally possible. It can't be done in Fortran 90 or C.

Also, of the memory management schemes that are lumped under the term of 'heap' ones, there are several that have deterministic behaviour and do not suffer from fragmentation. The simplest one is if all allocations are of the same size, and chains are used instead of arrays.

The whole area of memory management schemes is sadly neglected, and much of garbage collection work is based on unrealistic assumptions, so much of what is currently believed isn't so.

Regards, Nick Maclaren.

It *can* be done in C - by avoiding malloc ;).

Agreed. I've used techniques such as these (effectively an application-level memory manager) in certain situations where extreme reliability was required, but the platform insisted on the use of its memory manager... ...

Completely agreed. I've often fantasized about a hardware memory manager that remaps blocks in such a way that fragmentation cannot occur. OTOH, I've also fantasized that one day, a majority will write good code that doesn't leak memory and traps malloc errors properly...

Steve

That's a worthlessly vague statement. As far as I can see, the problem isn't lack of research; it's lack of application of that research in the most commonly used systems.

I will skip the mode nesting, as things are getting ridiculous :-)

Problem one: in C, there is no memory management beyond stack scoped except by using malloc. Any algorithm that requires more than that can't be done if you don't use malloc.

Problem two: there are many library facilities that use malloc either explicitly or implicitly - including almost all I/O - you would have to avoid them, too.

Regards, Nick Maclaren.

On the contrary - one of the main reasons that it hasn't been applied is the impracticality of the research. That was the case in 1965, and is the case today. A fairly good rule when you find that real engineers (and there are some real software engineers, even today, and used to be more) are ignoring the wonderful results that come out of 'scientific' research is to ask whether the research is actually addressing the whole of the problem that the engineers are faced with.

Regards, Nick Maclaren.

I'm posting this from c.a.e - embedded (or OS-less) is my main area. One tends not to need such I/O functions therein.

You're right, though - it certainly pays to be aware of the platform requirements of any such library function.

Steve

Does every research paper on memory management describe something practical? No. Does the research that has been done on memory management collectively address the problems that arise in real systems in a practical way, for those engineers that are prepared to do a bit of searching to find the papers relevant to those problems? Yes, it does. If you think otherwise, then be specific about the problems you think are not addressed.

sbrk()?

Provide your own malloc() so the I/O that needs it will get it from your own memory pool. Since you are doing the I/O, you can provide optimal buffering.

And still less the internationalisation and wide character facilities :-)

My experience of that area is that the worst problem is finding out when a compiler has chosen to implement some basic facility (e.g. a mathematical function) in a way that drags in completely unrelated junk from the library. But I have been usually trying to used hosted compilers in an embedded fashion - I would expect compilers designed for embedded use to be better.

Regards, Nick Maclaren.

I have, many times. You ignored the issues every time.

Regards, Nick Maclaren.

Not in C - in fact, I can't think of any standard that it IS in, because it makes such extremely horrible assumptions about the implementation.

No. Replacing malloc is not supported in C, and often breaks the nastier vendor and third-party libraries. The reason is that the suite of malloc-related functions is not independent, and they often use non-standard interfaces.

Yes, I have done it, over many years - but it isn't something that I recommend to users. Or can :-(

Regards, Nick Maclaren.

There are still systems where you can reasonably do so - shipping more than one possible malloc to use is a good hint for example.

In article , Sander Vesik writes: |> In comp.arch Nick Maclaren wrote: |> > |> > No. Replacing malloc is not supported in C, and often breaks |> > the nastier vendor and third-party libraries. The reason is that |> > the suite of malloc-related functions is not independent, and they |> > often use non-standard interfaces. |> |> There are still systems where you can reasonably do so - shipping |> more than one possible malloc to use is a good hint for example.

It's a hint that the VENDOR can do it - but not that the application programmer can. Yes, an experienced hacker can, which was the basis of my remark that I can, but I can't advise users to.

Also, are you aware of how often third-party libraries demand a particular one of those mallocs, either in their release notes or because they fall over with others?

Regards, Nick Maclaren.