By using this site, you agree to our updated Privacy Policy and our Terms of Use. Manage your Cookies Settings.
434,969 Members | 2,360 Online
Bytes IT Community
+ Ask a Question
Need help? Post your question and get tips & solutions from a community of 434,969 IT Pros & Developers. It's quick & easy.

Making Fatal Hidden Assumptions

P: n/a
We often find hidden, and totally unnecessary, assumptions being
made in code. The following leans heavily on one particular
example, which happens to be in C. However similar things can (and
do) occur in any language.

These assumptions are generally made because of familiarity with
the language. As a non-code example, consider the idea that the
faulty code is written by blackguards bent on foulling the
language. The term blackguards is not in favor these days, and for
good reason. However, the older you are, the more likely you are
to have used it since childhood, and to use it again, barring
specific thought on the subject. The same type of thing applies to
writing code.

I hope, with this little monograph, to encourage people to examine
some hidden assumptions they are making in their code. As ever, in
dealing with C, the reference standard is the ISO C standard.
Versions can be found in text and pdf format, by searching for N869
and N1124. [1] The latter does not have a text version, but is
more up-to-date.

We will always have innocent appearing code with these kinds of
assumptions built-in. However it would be wise to annotate such
code to make the assumptions explicit, which can avoid a great deal
of agony when the code is reused under other systems.

In the following example, the code is as downloaded from the
referenced URL, and the comments are entirely mine, including the
'every 5' linenumber references.

/* Making fatal hidden assumptions */
/* Paul Hsiehs version of strlen.
http://www.azillionmonkeys.com/qed/asmexample.html

Some sneaky hidden assumptions here:
1. p = s - 1 is valid. Not guaranteed. Careless coding.
2. cast (int) p is meaningful. Not guaranteed.
3. Use of 2's complement arithmetic.
4. ints have no trap representations or hidden bits.
5. 4 == sizeof(int) && 8 == CHAR_BIT.
6. size_t is actually int.
7. sizeof(int) is a power of 2.
8. int alignment depends on a zeroed bit field.

Since strlen is normally supplied by the system, the system
designer can guarantee all but item 1. Otherwise this is
not portable. Item 1 can probably be beaten by suitable
code reorganization to avoid the initial p = s - 1. This
is a serious bug which, for example, can cause segfaults
on many systems. It is most likely to foul when (int)s
has the value 0, and is meaningful.

He fails to make the valid assumption: 1 == sizeof(char).
*/

#define hasNulByte(x) ((x - 0x01010101) & ~x & 0x80808080)
#define SW (sizeof (int) / sizeof (char))

int xstrlen (const char *s) {
const char *p; /* 5 */
int d;

p = s - 1;
do {
p++; /* 10 */
if ((((int) p) & (SW - 1)) == 0) {
do {
d = *((int *) p);
p += SW;
} while (!hasNulByte (d)); /* 15 */
p -= SW;
}
} while (*p != 0);
return p - s;
} /* 20 */

Let us start with line 1! The constants appear to require that
sizeof(int) be 4, and that CHAR_BIT be precisely 8. I haven't
really looked too closely, and it is possible that the ~x term
allows for larger sizeof(int), but nothing allows for larger
CHAR_BIT. A further hidden assumption is that there are no trap
values in the representation of an int. Its functioning is
doubtful when sizeof(int) is less that 4. At the least it will
force promotion to long, which will seriously affect the speed.

This is an ingenious and speedy way of detecting a zero byte within
an int, provided the preconditions are met. There is nothing wrong
with it, PROVIDED we know when it is valid.

In line 2 we have the confusing use of sizeof(char), which is 1 by
definition. This just serves to obscure the fact that SW is
actually sizeof(int) later. No hidden assumptions have been made
here, but the usage helps to conceal later assumptions.

Line 4. Since this is intended to replace the systems strlen()
function, it would seem advantageous to use the appropriate
signature for the function. In particular strlen returns a size_t,
not an int. size_t is always unsigned.

In line 8 we come to a biggie. The standard specifically does not
guarantee the action of a pointer below an object. The only real
purpose of this statement is to compensate for the initial
increment in line 10. This can be avoided by rearrangement of the
code, which will then let the routine function where the
assumptions are valid. This is the only real error in the code
that I see.

In line 11 we have several hidden assumptions. The first is that
the cast of a pointer to an int is valid. This is never
guaranteed. A pointer can be much larger than an int, and may have
all sorts of non-integer like information embedded, such as segment
id. If sizeof(int) is less than 4 the validity of this is even
less likely.

Then we come to the purpose of the statement, which is to discover
if the pointer is suitably aligned for an int. It does this by
bit-anding with SW-1, which is the concealed sizeof(int)-1. This
won't be very useful if sizeof(int) is, say, 3 or any other
non-poweroftwo. In addition, it assumes that an aligned pointer
will have those bits zero. While this last is very likely in
todays systems, it is still an assumption. The system designer is
entitled to assume this, but user code is not.

Line 13 again uses the unwarranted cast of a pointer to an int.
This enables the use of the already suspicious macro hasNulByte in
line 15.

If all these assumptions are correct, line 19 finally calculates a
pointer difference (which is valid, and of type size_t or ssize_t,
but will always fit into a size_t). It then does a concealed cast
of this into an int, which could cause undefined or implementation
defined behaviour if the value exceeds what will fit into an int.
This one is also unnecessary, since it is trivial to define the
return type as size_t and guarantee success.

I haven't even mentioned the assumption of 2's complement
arithmetic, which I believe to be embedded in the hasNulByte
macro. I haven't bothered to think this out.

Would you believe that so many hidden assumptions can be embedded
in such innocent looking code? The sneaky thing is that the code
appears trivially correct at first glance. This is the stuff that
Heisenbugs are made of. Yet use of such code is fairly safe if we
are aware of those hidden assumptions.

I have cross-posted this without setting follow-ups, because I
believe that discussion will be valid in all the newsgroups posted.

[1] The draft C standards can be found at:
<http://www.open-std.org/jtc1/sc22/wg14/www/docs/>

--
"If you want to post a followup via groups.google.com, don't use
the broken "Reply" link at the bottom of the article. Click on
"show options" at the top of the article, then click on the
"Reply" at the bottom of the article headers." - Keith Thompson
More details at: <http://cfaj.freeshell.org/google/>
Also see <http://www.safalra.com/special/googlegroupsreply/>

Mar 6 '06 #1
Share this Question
Share on Google+
351 Replies


P: n/a

CBFalconer wrote (in part):
In the following example, the code is as downloaded from the
referenced URL, and the comments are entirely mine, including the
'every 5' linenumber references.

/* Making fatal hidden assumptions */
/* Paul Hsiehs version of strlen.
http://www.azillionmonkeys.com/qed/asmexample.html

Some sneaky hidden assumptions here:
1. p = s - 1 is valid. Not guaranteed. Careless coding.
2. cast (int) p is meaningful. Not guaranteed.
3. Use of 2's complement arithmetic.
4. ints have no trap representations or hidden bits.
5. 4 == sizeof(int) && 8 == CHAR_BIT.
6. size_t is actually int.
7. sizeof(int) is a power of 2.
8. int alignment depends on a zeroed bit field.
...


None of these objections is warranted in the original context,
where the code is given as transliteration of some x86 assembly
language. In effect, the author is offering a function that
might be used as part of a C implementation, where none of the
usual portability considerations need apply. Your commentary
really should have included some statement along those lines.

Mar 7 '06 #2

P: n/a

CBFalconer wrote:
/* Making fatal hidden assumptions */
/* Paul Hsiehs version of strlen.
I haven't read your entire treatise, and I surely wouldn't want to
defend
Mr. Hsieh (whose code, he's led us to understand, was written by his
dog)
but I do have some comments.
#define SW (sizeof (int) / sizeof (char))
The fact that sizeof(char)==1 doesn't make this line bad *if* the
programmer feels that viewing SW as chars per int is important.
I don't defend this particular silliness, but using an expression for
a simple constant is often good if it makes the constant
self-documenting.
Some sneaky hidden assumptions here:
1. p = s - 1 is valid. Not guaranteed. Careless coding.


Mr. Hsieh immediately does p++ and his code will be correct if then
p == s. I don't question Chuck's argument, or whether the C standard
allows the C compiler to trash the hard disk when it sees p=s-1,
but I'm sincerely curious whether anyone knows of an *actual*
environment
where p == s will ever be false after (p = s-1; p++).

Many of us fell in love with C because, in practice, it is so much
simpler
and more deterministic than many languages. Many of the discussions in
comp.lang.c seem like they'd be better in a new newsgroup:
comp.lang.i'd_rather_be_a_lawyer

:-) :-)

James Dow Allen

Mar 7 '06 #3

P: n/a
CBFalconer wrote:
We often find hidden, and totally unnecessary, assumptions being
made in code. The following leans heavily on one particular
example, which happens to be in C. However similar things can (and
do) occur in any language.

These assumptions are generally made because of familiarity with
the language. As a non-code example, consider the idea that the
faulty code is written by blackguards bent on foulling the
language. The term blackguards is not in favor these days, and for
good reason. However, the older you are, the more likely you are
to have used it since childhood, and to use it again, barring
specific thought on the subject. The same type of thing applies to
writing code.

I hope, with this little monograph, to encourage people to examine
some hidden assumptions they are making in their code. As ever, in
dealing with C, the reference standard is the ISO C standard.
Versions can be found in text and pdf format, by searching for N869
and N1124. [1] The latter does not have a text version, but is
more up-to-date.

Getting people to think about their code is no bad idea. I've added a
couple of comments below...
We will always have innocent appearing code with these kinds of
assumptions built-in. However it would be wise to annotate such
code to make the assumptions explicit, which can avoid a great deal
of agony when the code is reused under other systems.

In the following example, the code is as downloaded from the
referenced URL, and the comments are entirely mine, including the
'every 5' linenumber references.

/* Making fatal hidden assumptions */
/* Paul Hsiehs version of strlen.
http://www.azillionmonkeys.com/qed/asmexample.html

Some sneaky hidden assumptions here:
1. p = s - 1 is valid. Not guaranteed. Careless coding.
Not guaranteed in what way? You are not guaranteed that p will be a
valid pointer, but you don't require it to be a valid pointer - all that
is required is that "p = s - 1" followed by "p++" leaves p equal to s.
I'm not good enough at the laws of C to tell you if this is valid, but
the laws of real life implementation say that it *is* valid on standard
2's complement cpus, *assuming* you do not have any sort of saturated
arithmetic.

In other words, the assumption is valid on most systems, but may be
risky on some DSPs or on dinosaurs.
2. cast (int) p is meaningful. Not guaranteed.
In what way could it not be meaningful here? In particular, the code is
correct even in a 32-bit int, 64-bit pointer environment.
3. Use of 2's complement arithmetic.
Again, this is valid on everything bar a few dinosaurs.
4. ints have no trap representations or hidden bits.
Ditto.
5. 4 == sizeof(int) && 8 == CHAR_BIT.
This is clearly an assumption that is not valid in general (although
perfectly acceptable in the context of this webpage, which assumes an
x86 in 32-bit mode).
6. size_t is actually int.
No, the assumption is that the ratio of two size_t items is compatible
with int. This may or may not be valid according to the laws of C.
7. sizeof(int) is a power of 2.
This is implied by (5), so it's not a new assumption.
8. int alignment depends on a zeroed bit field.

I'm not quite sure what you mean here, but I think this is implied by
(3) and (4).
Since strlen is normally supplied by the system, the system
designer can guarantee all but item 1. Otherwise this is
not portable. Item 1 can probably be beaten by suitable
code reorganization to avoid the initial p = s - 1. This
is a serious bug which, for example, can cause segfaults
on many systems. It is most likely to foul when (int)s
has the value 0, and is meaningful.
That's incorrect. No machine (that I can imagine) would segfault on p =
s - 1. It might segfault if p is then dereferenced, but it never is
(until it has been incremented again).

He fails to make the valid assumption: 1 == sizeof(char).
*/

Since strlen is supplied by the system, it is reasonable to make
assumptions about the system when writing it. It is *impossible* to
write good C code for a low-level routine like this without making
assumptions. The nearest you can get to a portable solution is a
horrible mess involving pre-processor directives to generate different
code depending on the target architecture. *All* the assumptions made
here are valid in that context - the code is good.
As a general point, however, it is important to comment such routines
with their assumptions, and possibly use pre-processor directives to
give an error if the assumptions are not valid. In particular, this
code assumes a 32-bit int, 8-bit char model, and an otherwise standard CPU.

<snip similar analysis>

Would you believe that so many hidden assumptions can be embedded
in such innocent looking code? The sneaky thing is that the code
appears trivially correct at first glance. This is the stuff that
Heisenbugs are made of. Yet use of such code is fairly safe if we
are aware of those hidden assumptions.

I have cross-posted this without setting follow-ups, because I
believe that discussion will be valid in all the newsgroups posted.

[1] The draft C standards can be found at:
<http://www.open-std.org/jtc1/sc22/wg14/www/docs/>


I'm following this in comp.arch.embedded, where we see a lot of
different target architectures, and where efficient code is perhaps more
important than in the world of "big" computers (at least, a higher
proportion of "big" computer programmers seem to forget about
efficiency...). There are two things about such target-specific
assumptions, in embedded programming - you should make such assumptions,
for the sake of good code, and you must know what they are, for the sake
of correct and re-usable code.
Mar 7 '06 #4

P: n/a
en******@yahoo.com wrote:
CBFalconer wrote (in part):
In the following example, the code is as downloaded from the
referenced URL, and the comments are entirely mine, including the
'every 5' linenumber references.

/* Making fatal hidden assumptions */
/* Paul Hsiehs version of strlen.
http://www.azillionmonkeys.com/qed/asmexample.html

Some sneaky hidden assumptions here:
1. p = s - 1 is valid. Not guaranteed. Careless coding.
2. cast (int) p is meaningful. Not guaranteed.
Mentioned by Paul.
3. Use of 2's complement arithmetic.
4. ints have no trap representations or hidden bits.
5. 4 == sizeof(int) && 8 == CHAR_BIT.
Paul says it wouldmentionedtaken care of.
6. size_t is actually int.
7. sizeof(int) is a power of 2.
8. int alignment depends on a zeroed bit field.
...


None of these objections is warranted in the original context,
where the code is given as transliteration of some x86 assembly
language. In effect, the author is offering a function that
might be used as part of a C implementation, where none of the
usual portability considerations need apply. Your commentary
really should have included some statement along those lines.


Much as I dislike some of what Paul posts, on that site Paul explicitly
states in the text that the code is non-portable and mentions as least a
couple of issues.

Personally, I would not expect a function documented as being
non-portable and optimised for a specific processor to specify every
single assumption.
--
Flash Gordon, living in interesting times.
Web site - http://home.flash-gordon.me.uk/
comp.lang.c posting guidelines and intro:
http://clc-wiki.net/wiki/Intro_to_clc
Mar 7 '06 #5

P: n/a
David Brown wrote:

CBFalconer wrote:

http://www.azillionmonkeys.com/qed/asmexample.html

Some sneaky hidden assumptions here:
1. p = s - 1 is valid. Not guaranteed. Careless coding.


Not guaranteed in what way? You are not guaranteed that p will be a
valid pointer, but you don't require it to be a valid pointer
- all that
is required is that "p = s - 1" followed by "p++" leaves p equal to s.
I'm not good enough at the laws of C to tell you if this is valid,


Merely subtracting 1 from s, renders the entire code undefined.
You're "off the map" as far as the laws of C are concerned.
On comp.lang.c,
we're mostly interested in what the laws of C *do* say
is guaranteed to work.

--
pete
Mar 7 '06 #6

P: n/a
James Dow Allen wrote:

CBFalconer wrote:

Some sneaky hidden assumptions here:
1. p = s - 1 is valid. Not guaranteed. Careless coding.


Mr. Hsieh immediately does p++ and his code will be correct if then
p == s. I don't question Chuck's argument, or whether the C standard
allows the C compiler to trash the hard disk when it sees p=s-1,
but I'm sincerely curious whether anyone knows of an *actual*
environment
where p == s will ever be false after (p = s-1; p++).

Many of us fell in love with C because, in practice, it is so much
simpler
and more deterministic than many languages.
Many of the discussions in
comp.lang.c seem like they'd be better in a new newsgroup:
comp.lang.i'd_rather_be_a_lawyer


If you don't want to know
whether or not (s - 1) causes undefined behavior,
then don't let it bother you.
You can write code any way you want.
But, if you want to discuss whether or not
(s - 1) causes undefined behavior, it does,
and this is the place to find out about it.

--
pete
Mar 7 '06 #7

P: n/a
James Dow Allen said:

CBFalconer wrote:
#define SW (sizeof (int) / sizeof (char))
The fact that sizeof(char)==1 doesn't make this line bad *if* the
programmer feels that viewing SW as chars per int is important.
I don't defend this particular silliness, but using an expression for
a simple constant is often good if it makes the constant
self-documenting.


Specifically, there is a good argument for char *p = malloc(n * sizeof *p).
Some sneaky hidden assumptions here:
1. p = s - 1 is valid. Not guaranteed. Careless coding.
Mr. Hsieh immediately does p++ and his code will be correct if then
p == s.


Not so, since forming an invalid pointer invokes undefined behaviour.
I don't question Chuck's argument, or whether the C standard
allows the C compiler to trash the hard disk when it sees p=s-1,
It does.
but I'm sincerely curious whether anyone knows of an *actual*
environment
where p == s will ever be false after (p = s-1; p++).


Since it won't compile, the question is academic. But assuming you meant to
have a test in there somewhere, my personal answer would be: no, I don't
know of such an environment off the top of my head, but there are some
weird environments out there (the comp.lang.c FAQ lists a few), and it
would certainly not surprise me to learn of such an environment. In any
case, such an environment could become mainstream next week, or next year,
or next decade. Carefully-written code that observes the rules will
continue to work in such environments.

--
Richard Heathfield
"Usenet is a strange place" - dmr 29/7/1999
http://www.cpax.org.uk
email: rjh at above domain (but drop the www, obviously)
Mar 7 '06 #8

P: n/a
In article <44***************@yahoo.com>, cb********@yahoo.com says...
These assumptions are generally made because of familiarity with
the language. As a non-code example, consider the idea that the
faulty code is written by blackguards bent on foulling the
language. The term blackguards is not in favor these days, and for
good reason.


About as good a reason as the term niggardly, as far as I can tell.
Perhaps the words are appropriate in a post relating to fatal
assumptions.

- Gerry Quinn
Mar 7 '06 #9

P: n/a
On Tue, 07 Mar 2006 10:31:09 +0000, pete wrote:
David Brown wrote:

CBFalconer wrote:

> http://www.azillionmonkeys.com/qed/asmexample.html
>
> Some sneaky hidden assumptions here:
> 1. p = s - 1 is valid. Not guaranteed. Careless coding.


Not guaranteed in what way? You are not guaranteed that p will be a
valid pointer, but you don't require it to be a valid pointer - all that
is required is that "p = s - 1" followed by "p++" leaves p equal to s.
I'm not good enough at the laws of C to tell you if this is valid,


Merely subtracting 1 from s, renders the entire code undefined. You're
"off the map" as far as the laws of C are concerned. On comp.lang.c,
we're mostly interested in what the laws of C *do* say is guaranteed to
work.


It seems to me ironic that, in a discussion about hidden assumptions, the
truth of this remark requires a hidden assumption about how the function
is called. Unless I am missing something big, p = s - 1 is fine unless s
points to the first element of an array (or worse)[1]. One can imagine
situations such as "string pool" implementations where all strings are
part of the same array where p = s - 1 is well defined. (You'd have to
take care with the first string: char pool[BIG_NUMBER]; char *pool_start =
pool + 1;).

[1] Since this is a language law discussion, we need to take the
definition from the standard where T *s is deemed to point to an array of
size one if it points to an object of type T. By "worse" I mean that s
does not point into (or just past) an array at all.

--
Ben.
Mar 7 '06 #10

P: n/a
Ben Bacarisse wrote:
p = s - 1 is fine unless s
points to the first element of an array (or worse)[1].


My simple mind must be missing something big here. If for pointer p,
(p-1) is deprecated because it's not guaranteed that it points to
anything sensible, why is p++ OK? There's no boundary checking in C
(unless you put it in).

Quote from p98 of K&R 2nd (ANSI) edition: "If pa points to a particular
element of an array, then by definition pa+1 points to the next element,
pa+i points to i elements after pa, and pa-i points to i elements before."

Paul Burke
Mar 7 '06 #11

P: n/a

On Tue, 7 Mar 2006, Paul Burke wrote:
Ben Bacarisse wrote:

p = s - 1 is fine unless s
points to the first element of an array (or worse)[1].
My simple mind must be missing something big here. If for pointer p, (p-1)
is deprecated because it's not guaranteed that it points to anything
sensible, why is p++ OK? There's no boundary checking in C (unless you
put it in).


(BTW, "deprecated," in the context of programming-language standards,
means something that once was okay but is not recommended anymore. (p-1)
isn't like that.) Read on for your answer.
Quote from p98 of K&R 2nd (ANSI) edition: "If pa points to a particular
element of an array, then by definition pa+1 points to the next element,
pa+i points to i elements after pa, and pa-i points to i elements before."


K&R answers your question. If pa points to some element of an array,
then pa-1 points to the /previous element/. But what's the "previous
element" relative to the first element in the array? It doesn't exist.
So we have undefined behavior.
The expression pa+1 is similar, but with one special case. If pa points
to the last element in the array, you might expect that pa+1 would be
undefined; but actually the C standard specifically allows you to evaluate
pa+1 in that case. Dereferencing that pointer, or incrementing it /again/,
however, invoke undefined behavior.

Basically: A C pointer must always point to something. "The
negative-oneth element of array a" is not "something."

int a[10];
int *pa = a+3;
pa-3; /* fine; points to a[0] */
pa+6; /* fine; points to a[9] */
pa+7; /* fine; points "after" a[9] */
pa+8; /* undefined behavior */
pa-4; /* undefined behavior */

HTH,
-Arthur
Mar 7 '06 #12

P: n/a
Paul Burke <pa**@scazon.com> writes:
My simple mind must be missing something big here. If for pointer p,
(p-1) is deprecated because it's not guaranteed that it points to
anything sensible, why is p++ OK? There's no boundary checking in C
(unless you put it in).


You're missing what the standard says about it:

8 When an expression that has integer type is added to or
subtracted from a pointer, the result has the type of the
pointer operand. If the pointer operand points to an element
of an array object, and the array is large enough, the
result points to an element offset from the original element
such that the difference of the subscripts of the resulting
and original array elements equals the integer expression.
In other words, if the expression P points to the i-th
element of an array object, the expressions (P)+N
(equivalently, N+(P)) and (P)-N (where N has the value n)
point to, respectively, the i+n-th and i-n-th elements of
the array object, provided they exist. Moreover, if the
expression P points to the last element of an array object,
the expression (P)+1 points one past the last element of the
array object, and if the expression Q points one past the
last element of an array object, the expression (Q)-1 points
to the last element of the array object. If both the pointer
operand and the result point to elements of the same array
object, or one past the last element of the array object,
the evaluation shall not produce an overflow; otherwise, the
behavior is undefined. If the result points one past the
last element of the array object, it shall not be used as
the operand of a unary * operator that is evaluated.

(Wow, that's a mouthful.)
--
"To get the best out of this book, I strongly recommend that you read it."
--Richard Heathfield
Mar 7 '06 #13

P: n/a


Paul Burke wrote On 03/07/06 11:45,:
Ben Bacarisse wrote:

p = s - 1 is fine unless s
points to the first element of an array (or worse)[1].

My simple mind must be missing something big here. If for pointer p,
(p-1) is deprecated because it's not guaranteed that it points to
anything sensible, why is p++ OK? There's no boundary checking in C
(unless you put it in).

Quote from p98 of K&R 2nd (ANSI) edition: "If pa points to a particular
element of an array, then by definition pa+1 points to the next element,
pa+i points to i elements after pa, and pa-i points to i elements before."


There's a special dispensation that allows you to
calculate a pointer value that points to the imaginary
element "one position past the end" of an array. (You
cannot, of course, actually try to access the imaginary
element -- but you can form the pointer value, compare
it to other pointer values, and so on.)

There is no such special dispensation for an element
"one position before the beginning."

The Rationale says the Committee considered defining
the effects at both ends (bilateral dispensations?), but
rejected it for efficiency reasons. Consider an array of
large elements -- structs of 32KB size, say. A system
that actually performed hardware checking of pointer values
could accommodate the one-past-the-end rule by allocating
just one extra byte after the end of the array, a byte that
the special pointer value could point at without setting off
the hardware's alarms. But one-before-the-beginning would
require an extra 32KB, just to hold data that could never
be used ...

--
Er*********@sun.com

Mar 7 '06 #14

P: n/a
On Tue, 07 Mar 2006 16:45:06 +0000, Paul Burke wrote:
Ben Bacarisse wrote:
p = s - 1 is fine unless s
points to the first element of an array (or worse)[1].


My simple mind must be missing something big here. If for pointer p, (p-1)
is deprecated because it's not guaranteed that it points to anything
sensible, why is p++ OK? There's no boundary checking in C (unless you put
it in).


As your (snipped) quote from K&R illustrates, pointer arithmetic is
defined only "within" arrays. "Within", includes a pointer that points
just after the last element. This is very handy, since pointer
expressions like end - start are a common idiom. When a pointer, p,
points to the first element of an array, p - 1 is not defined (indeed it
may not be representable given a devious enough, but conforming,
implementation). p - 1 is not deprecated at all (so far as I know). It
is either perfectly valid or entirely undefined. p + 1 is well-defined if
p points to any array element (including the last). p + 2 is not defined
if p points to the last element of an array.

For the purpose of arithmetic, pointers to single data objects are treated
like pointers to arrays of size 1. So we have:

int x, a[2];

int *p = &x + 1; /* defined */
int *q = &x - 1; /* undefined */
int *r = &x + 2; /* undefined */

int *s = a; /* defined */
int *t = a - 1; /* undefined */
int *u = a + 1; /* defined */
int *v = a + 2; /* defined */
int *w = a + 3; /* undefined */

I hope I have not missed what you were talking about.

--
Ben.
Mar 7 '06 #15

P: n/a

"Ben Pfaff" <bl*@cs.stanford.edu> wrote in message
news:87************@benpfaff.org...
Paul Burke <pa**@scazon.com> writes:
My simple mind must be missing something big here. If for pointer p,
(p-1) is deprecated because it's not guaranteed that it points to
anything sensible, why is p++ OK? There's no boundary checking in C
(unless you put it in).


It seems I too have a simple mind. I read the recent replies to this and
found myself not sure I am better off.

This is what I think I understand:

int x;
int *p;
int *q;

p = &x; /* is OK */
p = &x + 1; /* is OK even though we have no idea what p points to */
p = &x + 6; /* is undefined - does this mean that p may not be the */
/* address six locations beyond x? */
/* or just that we don't know what is there? */
p = &x - 1; / as previous */

But the poster was comparing with p++, so...

p = &x; /* so far so good */
p++; /* still ok (?) but we dont know what is there */
p++; /* is this now undefined? */

I guess _my_ question is - in this context does 'undefined' mean just that
we cannot say anything about what the pointer points to or that we cannot
say anything about the value of the pointer. So for example:

p = &x;
q = &x;
p = p+8;
q = q+8;

should p and q have the same value or is that undefined.

Pete Harrison
Mar 7 '06 #16

P: n/a
In article <Pi**********************************@unix43.andre w.cmu.edu>,
Arthur J. O'Dwyer <aj*******@andrew.cmu.edu> wrote:
K&R answers your question. If pa points to some element of an array,
then pa-1 points to the /previous element/. But what's the "previous
element" relative to the first element in the array? It doesn't exist.
So we have undefined behavior.
The expression pa+1 is similar, but with one special case. If pa points
to the last element in the array, you might expect that pa+1 would be
undefined; but actually the C standard specifically allows you to evaluate
pa+1 in that case. Dereferencing that pointer, or incrementing it /again/,
however, invoke undefined behavior.


Right.

These limitations may perhaps be most easily be understood with
reference to the VAX "descriptor" model of pointers, in which a pointer
did not refer -directly- to the target memory, but instead refered to
a a -description- of that memory, including sizes and basic types
and current offset. In that scheme, with pa a pointer, pa-1 involves
internal work to produce the descriptor with the appropriate sizes and
offsets -- and at the time the relative offset was calculated, it would
be compared to the known bounds, and an exception could occur *then*
[when the pointer was built] rather than at the time the pointer was used.
Other circumstances where it could make a difference include cases in
which pa points to a block of memory at the very beginning of a memory
segment (on a segmented architecture). The calculation of the value of
pa-1 could then proceed in several ways:

- by exception (because the system notes the attempt to point before
the segment beginning)

- by wrapping the relative segment offset field to its maximum value
(which might or might not trigger odd behaviours)

- by wrapping the relative segment offset field to its maximum value -and-
decrementing the field that holds the address register number that
holds the base virtual memory (this effectively pointing into
a completely -different- block of memory, which might or might not
trigger odd behaviours)

- by noticing that the segment descriptor is not suitable for the
pointer and building a new segment descriptor that covers the extended
range (which might use up scarce segment descriptors unnecessarily)

- by exception (because the system notes that the new memory
address is not one that the user has access rights to)
There are undoubtedly other situations, but the point remains that
creating a pointer to "before" an object is not certain to work,
even if that pointer is never dereferenced.
--
Okay, buzzwords only. Two syllables, tops. -- Laurie Anderson
Mar 7 '06 #17

P: n/a

"Peter Harrison" <pe***@cannock.ac.uk> wrote in message
news:du**********@south.jnrs.ja.net...

"Ben Pfaff" <bl*@cs.stanford.edu> wrote in message
news:87************@benpfaff.org...
Paul Burke <pa**@scazon.com> writes:
My simple mind must be missing something big here. If for pointer p,
(p-1) is deprecated because it's not guaranteed that it points to
anything sensible, why is p++ OK? There's no boundary checking in C
(unless you put it in).


It seems I too have a simple mind. I read the recent replies to this and
found myself not sure I am better off.

This is what I think I understand:

int x;
int *p;
int *q;

p = &x; /* is OK */
p = &x + 1; /* is OK even though we have no idea what p points to */
p = &x + 6; /* is undefined - does this mean that p may not be the */
/* address six locations beyond x? */
/* or just that we don't know what is there? */
p = &x - 1; / as previous */

But the poster was comparing with p++, so...

p = &x; /* so far so good */
p++; /* still ok (?) but we dont know what is there */
p++; /* is this now undefined? */

I guess _my_ question is - in this context does 'undefined' mean just that
we cannot say anything about what the pointer points to or that we cannot
say anything about the value of the pointer. So for example:

p = &x;
q = &x;
p = p+8;
q = q+8;

should p and q have the same value or is that undefined.

Basically you're supposed to assume that the arithmetic can't be done if the
result (or an intermediate value) goes out of range. As if it were a
trapped overflow. The behaviour is undefined as soon as you evaluate the
r.h.s, so you don't even know that the assignment to p or q actually takes
place.

--
RSH


Mar 7 '06 #18

P: n/a
"Peter Harrison" <pe***@cannock.ac.uk> writes:
[...]
This is what I think I understand:

int x;
int *p;
int *q;

p = &x; /* is OK */
p = &x + 1; /* is OK even though we have no idea what p points to */
p = &x + 6; /* is undefined - does this mean that p may not be the */
/* address six locations beyond x? */
/* or just that we don't know what is there? */ [...] p = &x - 1; / as previous */

But the poster was comparing with p++, so...

p = &x; /* so far so good */
p++; /* still ok (?) but we dont know what is there */
p++; /* is this now undefined? */

I guess _my_ question is - in this context does 'undefined' mean just that
we cannot say anything about what the pointer points to or that we cannot
say anything about the value of the pointer. So for example:

p = &x;
q = &x;
p = p+8;
q = q+8;

should p and q have the same value or is that undefined.


"Undefined" means far more than that. It's not an undefined *value*,
it's undefined *behavior*, which the standard defines as:

behavior, upon use of a nonportable or erroneous program construct
or of erroneous data, for which this International Standard
imposes no requirements

NOTE Possible undefined behavior ranges from ignoring the
situation completely with unpredictable results, to behaving
during translation or program execution in a documented manner
characteristic of the environment (with or without the issuance of
a diagnostic message), to terminating a translation or execution
(with the issuance of a diagnostic message).

Once undefined behavior has occurred, it's meaningless to talk about
the value of p. If your program or your computer crashes, p doesn't
have a value; if demons start flying out of your nose (in accordance
with the standard joke around here), you'll have more important things
to worry about.

--
Keith Thompson (The_Other_Keith) ks***@mib.org <http://www.ghoti.net/~kst>
San Diego Supercomputer Center <*> <http://users.sdsc.edu/~kst>
We must do something. This is something. Therefore, we must do this.
Mar 7 '06 #19

P: n/a
On Tue, 07 Mar 2006 10:21:21 -0800, Ben Pfaff wrote:
Paul Burke <pa**@scazon.com> writes:
My simple mind must be missing something big here. If for pointer p,
(p-1) is deprecated because it's not guaranteed that it points to
anything sensible, why is p++ OK? There's no boundary checking in C
(unless you put it in).


You're missing what the standard says about it:

8 When an expression that has integer type is added to or
subtracted from a pointer, the result has the type of the pointer
operand. If the pointer operand points to an element of an array
object, and the array is large enough, the result points to an
element offset from the original element such that the difference of
the subscripts of the resulting and original array elements equals
the integer expression. In other words, if the expression P points to
the i-th element of an array object, the expressions (P)+N
(equivalently, N+(P)) and (P)-N (where N has the value n) point to,
respectively, the i+n-th and i-n-th elements of the array object,
provided they exist. Moreover, if the expression P points to the
last element of an array object, the expression (P)+1 points one past
the last element of the array object, and if the expression Q points
one past the last element of an array object, the expression (Q)-1
points to the last element of the array object. If both the pointer
operand and the result point to elements of the same array object, or
one past the last element of the array object, the evaluation shall
not produce an overflow; otherwise, the behavior is undefined. If the
result points one past the last element of the array object, it shall
not be used as the operand of a unary * operator that is evaluated.

(Wow, that's a mouthful.)


It's precicely this sort of tomfoolery on the part of the C standards
committee that has brought the language into such ill-repute in recent
years. It's practically unworkable now, compared to how it was in (say)
the immediately post-ANSI-fication years.

The code in question could trivially have p replaced by (p+p_index)
everywhere, where p_index is an int, and all of the arithmetic currently
effected on p is instead effected on p_index. I.e., if p was set to s,
and p_index set to -1, and p_index++ appeared as the first element inside
the outer do loop.

So having the standard make the semantic equivalent "undefined" only
serves to make the standard itself ever more pointless.

Bah, humbug. Think I'll go back to assembly language, where pointers do
what you tell them to, and don't complain about it to their lawyers.

--
Andrew

Mar 7 '06 #20

P: n/a
On Tue, 07 Mar 2006 13:28:37 -0500, Arthur J. O'Dwyer wrote:
K&R answers your question. If pa points to some element of an array,
then pa-1 points to the /previous element/. But what's the "previous
element" relative to the first element in the array? It doesn't exist. So
we have undefined behavior.
Only because the standard says so. Didn't have to be that way. There are
plenty of logically correct algorithms that could exist that involve
pointers that point somewhere outside of a[0..N]. As long as there's no
de-referencing, no harm, no foul. (Consider the simple case of iterating
through an array at a non-unit stride, using the normal p < s + N
termination condition. The loop finishes with p > s + N and the standard
says "pow, you're dead", when the semantically identical code written with
integer indexes has no legal problems.
The expression pa+1 is similar, but with one special case. If pa
points
to the last element in the array, you might expect that pa+1 would be
undefined; but actually the C standard specifically allows you to
evaluate pa+1 in that case. Dereferencing that pointer, or incrementing
it /again/, however, invoke undefined behavior.

Basically: A C pointer must always point to something. "The
negative-oneth element of array a" is not "something."


Only because the standard says so. The standard is stupid, in that
respect.

--
Andrew

Mar 7 '06 #21

P: n/a
On Tue, 07 Mar 2006 13:29:11 -0500, Eric Sosman wrote:
The Rationale says the Committee considered defining
the effects at both ends (bilateral dispensations?), but rejected it for
efficiency reasons. Consider an array of large elements -- structs of
32KB size, say. A system that actually performed hardware checking of
pointer values could accommodate the one-past-the-end rule by allocating
just one extra byte after the end of the array, a byte that the special
pointer value could point at without setting off the hardware's alarms.
But one-before-the-beginning would require an extra 32KB, just to hold
data that could never be used ...


So the standards body broke decades of practice and perfectly safe and
reasonable code to support a *hypothetical* implementation that was so
stupid that it checked pointer values, rather than pointer *use*?
Amazing.

--
Andrew

Mar 7 '06 #22

P: n/a
Andrew Reilly <an*************@areilly.bpc-users.org> writes:
It's precicely this sort of tomfoolery on the part of the C standards
committee that has brought the language into such ill-repute in recent
years. It's practically unworkable now, compared to how it was in (say)
the immediately post-ANSI-fication years.


The semantics you're complaining (one-past-the-end) about didn't
change in important ways from C89 to C99. I don't know why you'd
think the C99 semantics are unreasonable if you didn't think the
C89 semantics were unreasonable.
--
int main(void){char p[]="ABCDEFGHIJKLMNOPQRSTUVWXYZabcdefghijklmnopqrstuv wxyz.\
\n",*q="kl BIcNBFr.NKEzjwCIxNJC";int i=sizeof p/2;char *strchr();int putchar(\
);while(*q){i+=strchr(p,*q++)-p;if(i>=(int)sizeof p)i-=sizeof p-1;putchar(p[i]\
);}return 0;}
Mar 7 '06 #23

P: n/a
Andrew Reilly <an*************@areilly.bpc-users.org> writes:
On Tue, 07 Mar 2006 13:28:37 -0500, Arthur J. O'Dwyer wrote:
K&R answers your question. If pa points to some element of an array,
then pa-1 points to the /previous element/. But what's the "previous
element" relative to the first element in the array? It doesn't exist. So
we have undefined behavior.


Only because the standard says so. Didn't have to be that way. There are
plenty of logically correct algorithms that could exist that involve
pointers that point somewhere outside of a[0..N]. As long as there's no
de-referencing, no harm, no foul. (Consider the simple case of iterating
through an array at a non-unit stride, using the normal p < s + N
termination condition. The loop finishes with p > s + N and the standard
says "pow, you're dead", when the semantically identical code written with
integer indexes has no legal problems.


The standard is specifically designed to allow for architectures where
constructing an invalid pointer value can cause a trap even if the
pointer is not dereferenced.

For example, given:

int n;
int *ptr1 = &n; /* ptr1 points to n */
int *ptr2 = ptr1 - 1 /* ptr2 points *before* n in memory; this
invokes undefined behavior. */

Suppose some CPU uses special instructions and registers for pointer
values. Suppose arr happens to be allocated at the very beginning of
a memory segment. Just constructing the value ptr1-1 could cause a
trap of some sort. Or it could quietly yield a value such that
ptr2+1 != ptr1.

By saying that this is undefined behavior, the C standard isn't
forbidding you to do it; it's just refusing to tell you how it
behaves. If you're using an implementation that guarantees that this
will work the way you want it to, and if you're not concerned about
portability to other implementations, there's nothing stopping you
from doing it.

On the other hand, if the standard had defined the behavior of this
construct, it would have required all implementations to support it.
On a system that does strong checking of all pointer values, a C
compiler might have to generate inefficient code to meet the
standard's requirements.

--
Keith Thompson (The_Other_Keith) ks***@mib.org <http://www.ghoti.net/~kst>
San Diego Supercomputer Center <*> <http://users.sdsc.edu/~kst>
We must do something. This is something. Therefore, we must do this.
Mar 7 '06 #24

P: n/a
Andrew Reilly <an*************@areilly.bpc-users.org> writes:
So the standards body broke decades of practice and perfectly safe and
reasonable code to support a *hypothetical* implementation that was so
stupid that it checked pointer values, rather than pointer *use*?


You can still write your code to make whatever assumptions you
like. You just can't assume that it will work portably. If, for
example, you are writing code for a particular embedded
architecture with a given compiler, then it may be reasonable to
make assumptions beyond those granted by the standard.

In other words, the standard provides minimum guarantees. Your
implementation may provide stronger ones.
--
"It would be a much better example of undefined behavior
if the behavior were undefined."
--Michael Rubenstein
Mar 7 '06 #25

P: n/a
On Tue, 07 Mar 2006 22:26:57 +0000, Keith Thompson wrote:
Andrew Reilly <an*************@areilly.bpc-users.org> writes:
On Tue, 07 Mar 2006 13:28:37 -0500, Arthur J. O'Dwyer wrote:
K&R answers your question. If pa points to some element of an array,
then pa-1 points to the /previous element/. But what's the "previous
element" relative to the first element in the array? It doesn't exist.
So we have undefined behavior.
Only because the standard says so. Didn't have to be that way. There
are plenty of logically correct algorithms that could exist that involve
pointers that point somewhere outside of a[0..N]. As long as there's no
de-referencing, no harm, no foul. (Consider the simple case of
iterating through an array at a non-unit stride, using the normal p < s
+ N termination condition. The loop finishes with p > s + N and the
standard says "pow, you're dead", when the semantically identical code
written with integer indexes has no legal problems.


The standard is specifically designed to allow for architectures where
constructing an invalid pointer value can cause a trap even if the pointer
is not dereferenced.


And are there any? Any in common use? Any where the equivalent (well
defined) pointer+offset code would be slower?

I'll need that list to know which suppliers to avoid...
For example, given:

int n;
int *ptr1 = &n; /* ptr1 points to n */ int *ptr2 = ptr1 - 1 /*
ptr2 points *before* n in memory; this
invokes undefined behavior. */

Suppose some CPU uses special instructions and registers for pointer
values. Suppose arr happens to be allocated at the very beginning of a
memory segment. Just constructing the value ptr1-1 could cause a trap
of some sort. Or it could quietly yield a value such that ptr2+1 !=
ptr1.
Suppose the computer uses tribits.

Standards are meant to codify common practice. If you want a language
that only has object references and array indices, there are plenty of
those to chose from.
By saying that this is undefined behavior, the C standard isn't
forbidding you to do it; it's just refusing to tell you how it behaves.
And that helps who?
If you're using an implementation that guarantees that this will work
the way you want it to, and if you're not concerned about portability to
other implementations, there's nothing stopping you from doing it.
Which implementations?
On the other hand, if the standard had defined the behavior of this
construct, it would have required all implementations to support it. On
a system that does strong checking of all pointer values, a C compiler
might have to generate inefficient code to meet the standard's
requirements.


That would be a *good* thing. Checking any earlier than at reference time
breaks what it is about C that makes it C.

OK, I've made enough of a fool of myself already. I'll go and have that
second cup of coffee for the morning, before I start going on about having
the standard support non-2's complement integers, or machines that have no
arithmetic right shifts...

--
Andrew

Mar 7 '06 #26

P: n/a
On 2006-03-07, Andrew Reilly <an*************@areilly.bpc-users.org> wrote:
OK, I've made enough of a fool of myself already. I'll go and have that
second cup of coffee for the morning, before I start going on about having
the standard support non-2's complement integers, or machines that have no
arithmetic right shifts...


Arithmetic right shift isn't particularly useful on some machines that
do. notably twos-complement ones.
Mar 7 '06 #27

P: n/a
On Wed, 08 Mar 2006 07:57:39 +1100, Andrew Reilly
<an*************@areilly.bpc-users.org> wrote:
It's precicely this sort of tomfoolery on the part of the C standards
committee that has brought the language into such ill-repute in recent
years. It's practically unworkable now, compared to how it was in (say)
the immediately post-ANSI-fication years.


Well, shucks, I manage to make it work pretty well most every day.
Does that mean I'm in ill-repute too?

--
Al Balmer
Sun City, AZ
Mar 7 '06 #28

P: n/a
On Wed, 08 Mar 2006 08:14:37 +1100, Andrew Reilly
<an*************@areilly.bpc-users.org> wrote:
On Tue, 07 Mar 2006 13:29:11 -0500, Eric Sosman wrote:
The Rationale says the Committee considered defining
the effects at both ends (bilateral dispensations?), but rejected it for
efficiency reasons. Consider an array of large elements -- structs of
32KB size, say. A system that actually performed hardware checking of
pointer values could accommodate the one-past-the-end rule by allocating
just one extra byte after the end of the array, a byte that the special
pointer value could point at without setting off the hardware's alarms.
But one-before-the-beginning would require an extra 32KB, just to hold
data that could never be used ...


So the standards body broke decades of practice and perfectly safe and
reasonable code to support a *hypothetical* implementation that was so
stupid that it checked pointer values, rather than pointer *use*?
Amazing.


Decades of use? This isn't a new rule.

An implementation might choose, for valid reasons, to prefetch the
data that pointer is pointing to. If it's in a segment not allocated
....

--
Al Balmer
Sun City, AZ
Mar 7 '06 #29

P: n/a
Andrew Reilly <an*************@areilly.bpc-users.org> writes:
On Tue, 07 Mar 2006 22:26:57 +0000, Keith Thompson wrote:
The standard is specifically designed to allow for architectures where
constructing an invalid pointer value can cause a trap even if the pointer
is not dereferenced.


And are there any? Any in common use?


x86 in non-flat protected mode would be one example. Attempting
to load an invalid value into a segment register causes a fault.
--
"Given that computing power increases exponentially with time,
algorithms with exponential or better O-notations
are actually linear with a large constant."
--Mike Lee
Mar 7 '06 #30

P: n/a
On Tue, 07 Mar 2006 14:54:30 -0800, Ben Pfaff wrote:
Andrew Reilly <an*************@areilly.bpc-users.org> writes:
On Tue, 07 Mar 2006 22:26:57 +0000, Keith Thompson wrote:
The standard is specifically designed to allow for architectures where
constructing an invalid pointer value can cause a trap even if the
pointer is not dereferenced.


And are there any? Any in common use?


x86 in non-flat protected mode would be one example. Attempting to load
an invalid value into a segment register causes a fault.


I wasn't aware that address arithmetic generally operated on the segment
register in that environment, rather on the "pointer" register used within
the segment. I haven't coded in that environment myself, so I have no
direct experience to call on. My understanding was that the architecture
was intrinsically a segment+offset mechanism, so having the compiler
produce the obvious code in the offset value (i.e., -1) would not incur
the performance penalty that has been mentioned. (Indeed, it's loading
the segment register that causes the performance penalty, I believe.)

--
Andrew

Mar 7 '06 #31

P: n/a
On Tue, 07 Mar 2006 15:59:37 -0700, Al Balmer wrote:
An implementation might choose, for valid reasons, to prefetch the data
that pointer is pointing to. If it's in a segment not allocated ...


Hypothetical hardware that traps on *speculative* loads isn't broken by
design? I'd love to see the initialization sequences, or the task
switching code that has to make sure that all pointer values are valid
before they're loaded. No, scratch that. I've got better things to do.

Cheers,

--
Andrew

Mar 7 '06 #32

P: n/a
Andrew Reilly <an*************@areilly.bpc-users.org> writes:
On Tue, 07 Mar 2006 22:26:57 +0000, Keith Thompson wrote:
Andrew Reilly <an*************@areilly.bpc-users.org> writes:
On Tue, 07 Mar 2006 13:28:37 -0500, Arthur J. O'Dwyer wrote:
K&R answers your question. If pa points to some element of an array,
then pa-1 points to the /previous element/. But what's the "previous
element" relative to the first element in the array? It doesn't exist.
So we have undefined behavior.

Only because the standard says so. Didn't have to be that way. There
are plenty of logically correct algorithms that could exist that involve
pointers that point somewhere outside of a[0..N]. As long as there's no
de-referencing, no harm, no foul. (Consider the simple case of
iterating through an array at a non-unit stride, using the normal p < s
+ N termination condition. The loop finishes with p > s + N and the
standard says "pow, you're dead", when the semantically identical code
written with integer indexes has no legal problems.
The standard is specifically designed to allow for architectures where
constructing an invalid pointer value can cause a trap even if the pointer
is not dereferenced.


And are there any? Any in common use? Any where the equivalent (well
defined) pointer+offset code would be slower?


I really don't know, but the idea of allowing errors to be caught as
early as possible seems like a good one.

[...]
Suppose the computer uses tribits.
Do you mean trinary digits rather than binary digits? The C standard
requires binary representation for integers.
Standards are meant to codify common practice. If you want a language
that only has object references and array indices, there are plenty of
those to chose from.
By saying that this is undefined behavior, the C standard isn't
forbidding you to do it; it's just refusing to tell you how it behaves.
And that helps who?


It (potentially) helps implementers to generate the most efficient
possible code, and it helps programmers to know what's actually
guaranteed to work across all possible platforms with conforming C
implementations.

[...]
OK, I've made enough of a fool of myself already. I'll go and have that
second cup of coffee for the morning, before I start going on about having
the standard support non-2's complement integers, or machines that have no
arithmetic right shifts...


C99 allows signed integers to be represented in 2's-complement,
1's-complement, or signed-magnitude (I think I mispunctuated at least
one of those).

C has been implemented on machines that don't support floating-point,
or even multiplication and division, in hardware. The compiler just
has to do whatever is necessary to meet the standard's requirements.

--
Keith Thompson (The_Other_Keith) ks***@mib.org <http://www.ghoti.net/~kst>
San Diego Supercomputer Center <*> <http://users.sdsc.edu/~kst>
We must do something. This is something. Therefore, we must do this.
Mar 7 '06 #33

P: n/a
Andrew Reilly <an*************@areilly.bpc-users.org> writes:
On Tue, 07 Mar 2006 14:54:30 -0800, Ben Pfaff wrote:
Andrew Reilly <an*************@areilly.bpc-users.org> writes:
On Tue, 07 Mar 2006 22:26:57 +0000, Keith Thompson wrote:

The standard is specifically designed to allow for architectures where
constructing an invalid pointer value can cause a trap even if the
pointer is not dereferenced.

And are there any? Any in common use?


x86 in non-flat protected mode would be one example. Attempting to load
an invalid value into a segment register causes a fault.


I wasn't aware that address arithmetic generally operated on the segment
register in that environment, rather on the "pointer" register used within
the segment. [...]


Address arithmetic might not, but the standard doesn't disallow
it. Other uses of invalid pointers, e.g. comparing a pointer
into a freed memory block against some other pointer, seem more
likely to do so.
--
"Give me a couple of years and a large research grant,
and I'll give you a receipt." --Richard Heathfield
Mar 7 '06 #34

P: n/a
Andrew Reilly <an*************@areilly.bpc-users.org> writes:
On Tue, 07 Mar 2006 13:29:11 -0500, Eric Sosman wrote:
The Rationale says the Committee considered defining
the effects at both ends (bilateral dispensations?), but rejected it for
efficiency reasons. Consider an array of large elements -- structs of
32KB size, say. A system that actually performed hardware checking of
pointer values could accommodate the one-past-the-end rule by allocating
just one extra byte after the end of the array, a byte that the special
pointer value could point at without setting off the hardware's alarms.
But one-before-the-beginning would require an extra 32KB, just to hold
data that could never be used ...


So the standards body broke decades of practice and perfectly safe and
reasonable code to support a *hypothetical* implementation that was so
stupid that it checked pointer values, rather than pointer *use*?
Amazing.


Consider the alternative.

#define LEN 100
#define INC 5000
int arr[LEN];
int *ptr = arr;
/* A */
ptr += 2*INC;
ptr -= INC;
ptr -= INC;
/* B */
ptr -= INC;
ptr -= INC;
ptr += 2*INC;
/* C */

What you're suggesting, I think, is that ptr==arr should be true at
points A, B, and C, for any(?) values of LEN and INC. It happens to
work out that way sometimes (at least in the test program I just
tried), but I can easily imagine a system where guaranteeing this
would place an undue burden on the compiler.

--
Keith Thompson (The_Other_Keith) ks***@mib.org <http://www.ghoti.net/~kst>
San Diego Supercomputer Center <*> <http://users.sdsc.edu/~kst>
We must do something. This is something. Therefore, we must do this.
Mar 7 '06 #35

P: n/a
On Wed, 08 Mar 2006 09:39:33 +1100, Andrew Reilly
<an*************@areilly.bpc-users.org> wrote:
On Tue, 07 Mar 2006 15:59:37 -0700, Al Balmer wrote:
An implementation might choose, for valid reasons, to prefetch the data
that pointer is pointing to. If it's in a segment not allocated ...


Hypothetical hardware that traps on *speculative* loads isn't broken by
design? I'd love to see the initialization sequences, or the task
switching code that has to make sure that all pointer values are valid
before they're loaded. No, scratch that. I've got better things to do.

It doesn't have to make sure. It's free to segfault. You write funny
code, you pay the penalty (or your customers do.) Modern hardware does
a lot of speculation. It can preload or even precompute both branches
of a conditional, for example.

--
Al Balmer
Sun City, AZ
Mar 7 '06 #36

P: n/a


Andrew Reilly wrote On 03/07/06 16:41,:
On Tue, 07 Mar 2006 22:26:57 +0000, Keith Thompson wrote:

The standard is specifically designed to allow for architectures where
constructing an invalid pointer value can cause a trap even if the pointer
is not dereferenced.


And are there any? Any in common use? Any where the equivalent (well
defined) pointer+offset code would be slower?


I've been told that IBM AS/400 bollixes the bogus
arithmetic, at least under some circumstances. A friend
told of fixing code that did something like

if (buffer_pos + amount_needed > buffer_limit) {
... enlarge the buffer ...
}
memcpy (buffer_pos, from_somewhere, amount_needed);
buffer_pos += amount_needed;

This looks innocuous to devotees of flat address spaces
(Flat-Earthers?), but it didn't work on AS/400. If the
sum `buffer_pos + amount_needed' went past the end of the
buffer, the result was some kind of NaP ("not a pointer")
and the comparison didn't kick in. Result: the code never
discovered that the buffer needed enlarging, and merrily
tried to memcpy() into a too-small area ...

I have no personal experience of the AS/400, and I may
have misremembered some of what my friend related. Would
anybody with AS/400 knowledge care to comment?

--
Er*********@sun.com

Mar 7 '06 #37

P: n/a
Andrew Reilly wrote
(in article
<pa****************************@areilly.bpc-users.org>):
On Tue, 07 Mar 2006 15:59:37 -0700, Al Balmer wrote:
An implementation might choose, for valid reasons, to prefetch the data
that pointer is pointing to. If it's in a segment not allocated ...


Hypothetical hardware that traps on *speculative* loads isn't broken by
design? I'd love to see the initialization sequences, or the task
switching code that has to make sure that all pointer values are valid
before they're loaded. No, scratch that. I've got better things to do.


This is a lot of whining about a specific problem that can
easily be remedied just by changing the loop construction. The
whole debate is pretty pointless in that context, unless you
have some religious reason to insist upon the method in the
original.
--
Randy Howard (2reply remove FOOBAR)
"The power of accurate observation is called cynicism by those
who have not got it." - George Bernard Shaw

Mar 7 '06 #38

P: n/a

"Andrew Reilly" <an*************@areilly.bpc-users.org> wrote in message
news:pa****************************@areilly.bpc-users.org...
On Tue, 07 Mar 2006 10:21:21 -0800, Ben Pfaff wrote:
Paul Burke <pa**@scazon.com> writes:
My simple mind must be missing something big here. If for pointer p,
(p-1) is deprecated because it's not guaranteed that it points to
anything sensible, why is p++ OK? There's no boundary checking in C
(unless you put it in).
You're missing what the standard says about it:

8 When an expression that has integer type is added to or
subtracted from a pointer, the result has the type of the pointer
operand. If the pointer operand points to an element of an array
object, and the array is large enough, the result points to an
element offset from the original element such that the difference of the subscripts of the resulting and original array elements equals
the integer expression. In other words, if the expression P points to the i-th element of an array object, the expressions (P)+N
(equivalently, N+(P)) and (P)-N (where N has the value n) point to,
respectively, the i+n-th and i-n-th elements of the array object,
provided they exist. Moreover, if the expression P points to the
last element of an array object, the expression (P)+1 points one past the last element of the array object, and if the expression Q points one past the last element of an array object, the expression (Q)-1
points to the last element of the array object. If both the pointer
operand and the result point to elements of the same array object, or one past the last element of the array object, the evaluation shall
not produce an overflow; otherwise, the behavior is undefined. If the result points one past the last element of the array object, it shall not be used as the operand of a unary * operator that is evaluated.

(Wow, that's a mouthful.)


It's precicely this sort of tomfoolery on the part of the C standards
committee that has brought the language into such ill-repute in recent
years. It's practically unworkable now, compared to how it was in (say)
the immediately post-ANSI-fication years.

The code in question could trivially have p replaced by (p+p_index)
everywhere, where p_index is an int, and all of the arithmetic currently
effected on p is instead effected on p_index. I.e., if p was set to s,
and p_index set to -1, and p_index++ appeared as the first element inside
the outer do loop.

So having the standard make the semantic equivalent "undefined" only
serves to make the standard itself ever more pointless.


It's not always equivalent. The trouble starts with

char a[8];
char *p;

for ( p = a+1 ; p < a+8 ; p += 2 ) {}

intending that the loop terminates on p == a+9 (since it skips a+8). But
how do we know that a+9 > a+8 ? If the array is right at the top of some
kind of segment, the arithmetic might have wrapped round.

To support predictable pointer comparisons out of range, the compiler would
have to allocate space with a safe buffer zone. Complications are setting
in.

Ints have the nice property that 0 is in the middle and we know how much
headroom we've got either side. So it's easy for the compiler to make the
int version work (leaving it to the programmer to take responsibility for
avoiding overflow, which is no big deal).

Pointers don't have that property. The compiler can't take sole
responsibility for avoiding overflow irrespective of what the programmer
does. If the programmer wants to go out of range and is at the same time
responsible for avoiding overflow, then he has to start worrying about
whereabouts his object is and what headroom he's got.

Bah, humbug. Think I'll go back to assembly language, where pointers do
what you tell them to, and don't complain about it to their lawyers.


Can't see how assembly programmers avoid the same kind of issue. I can see
how they could ignore it. The code will work most of the time.

--
RSH
Mar 8 '06 #39

P: n/a
On Tue, 07 Mar 2006 23:57:38 +0000, Keith Thompson wrote:
Consider the alternative.

#define LEN 100
#define INC 5000
int arr[LEN];
int *ptr = arr;
/* A */
ptr += 2*INC;
ptr -= INC;
ptr -= INC;
/* B */
ptr -= INC;
ptr -= INC;
ptr += 2*INC;
/* C */

What you're suggesting, I think, is that ptr==arr should be true at points
A, B, and C, for any(?) values of LEN and INC. It happens to work out
that way sometimes (at least in the test program I just tried), but I can
easily imagine a system where guaranteeing this would place an undue
burden on the compiler.


That *is* what I'm suggesting. In fact, I'm suggesting that p += a; p -=
a; should leave p as it was originally for any int a and pointer p. To my
mind, and I've been using C for more than 20 years, that is the very
essence of the nature of C. It's what makes pointer-as-cursor algorithms
make sense. Throw it away, and you might as well restrict yourself to
coding p[a], and then you've got fortran, pascal or Java.

Just because hardware can be imagined (or even built) that doesn't match
the conventional processor model that C most naturally fits *shouldn't* be
an argument to dilute or mess around with the C spec. Just use a
different language on those processors, or put up with some inefficiency
or compiler switches. Pascal has always been a pretty nice fit for many
hardware-pointer-checked machines. Such hardware isn't even a good
argument in this case though, since the obvious implementation will
involve base+offset compound pointers anyway, and mucking around with the
offset (as an integer) should neither trap nor cause a performance issue.

I've coded for years on Motorola 56000-series DSPs, and they don't look
anything like the conventional processor that C knows about: you've got
two separate data memory spaces and a third for program memory, pointers
aren't integers, words are 24-bits long and that's the smallest
addressable unit, and so on. Never the less,
there have been at least two C compilers for the thing, and they've both
produced *awful* code, and that's OK: they were never used for
performance-critical code. That was always done in assembler. There are
lots of processors (particularly DSPs) that are worse. I know of one that
doesn't have pointers as such at all. That's OK too. There isn't a C
compiler for that.

C is useful, though, and there's a lot of code written in it, so it's no
surprise that most of the more recent DSP designs actually do fit nicely
into the C conventional machine model. And (p + n) - n works in the
obvious fashion for those, too.

Cheers,

--
Andrew

Mar 8 '06 #40

P: n/a
In article <44********@news.wineasy.se> David Brown <da***@westcontrol.removethisbit.com> writes:
CBFalconer wrote:

....
Some sneaky hidden assumptions here:
1. p = s - 1 is valid. Not guaranteed. Careless coding.


Not guaranteed in what way? You are not guaranteed that p will be a
valid pointer, but you don't require it to be a valid pointer - all that
is required is that "p = s - 1" followed by "p++" leaves p equal to s.


But the standard allows "p = s - 1" to trap when an invalid pointer is
generated. And this can indeed be the case on segmented architectures.
--
dik t. winter, cwi, kruislaan 413, 1098 sj amsterdam, nederland, +31205924131
home: bovenover 215, 1025 jn amsterdam, nederland; http://www.cwi.nl/~dik/
Mar 8 '06 #41

P: n/a
On Tue, 07 Mar 2006 19:22:57 -0500, Eric Sosman wrote:


Andrew Reilly wrote On 03/07/06 16:41,:
On Tue, 07 Mar 2006 22:26:57 +0000, Keith Thompson wrote:

The standard is specifically designed to allow for architectures where
constructing an invalid pointer value can cause a trap even if the
pointer is not dereferenced.


And are there any? Any in common use? Any where the equivalent (well
defined) pointer+offset code would be slower?


I've been told that IBM AS/400 bollixes the bogus
arithmetic, at least under some circumstances. A friend told of fixing
code that did something like

if (buffer_pos + amount_needed > buffer_limit) {
... enlarge the buffer ...
}
memcpy (buffer_pos, from_somewhere, amount_needed); buffer_pos +=
amount_needed;

This looks innocuous to devotees of flat address spaces (Flat-Earthers?),
but it didn't work on AS/400. If the sum `buffer_pos + amount_needed'
went past the end of the buffer, the result was some kind of NaP ("not a
pointer") and the comparison didn't kick in. Result: the code never
discovered that the buffer needed enlarging, and merrily tried to memcpy()
into a too-small area ...

I have no personal experience of the AS/400, and I may
have misremembered some of what my friend related. Would anybody with
AS/400 knowledge care to comment?


I haven't used an AS/400 myself, either, but this is almost certainly the
sort of perfectly reasonable code that the standard has arranged to be
undefined, precicely so that it can be said that there's a C compiler for
that system.

Given the hardware behaviour, it would have been vastly preferable for the
compiler to handle pointers as base+offset pairs, so that the specialness
of the hardware pointers didn't interfere with the logic of the program.

Since most coding for AS/400s were (is still?) done in COBOL and PL/1,
both of which are perfectly suited to the hardware's two-dimensional
memory, any performance degredation would hardly have been noticed. (And
since AS/400s are actually Power processors with a JIT over the top now,
there would likely not be a performance problem from doing it "right"
anyway.) But no, your friend had to go and modify good code, and risk
introducing bugs in the process.

--
Andrew

Mar 8 '06 #42

P: n/a

"Gerry Quinn" <ge****@DELETETHISindigo.ie> wrote in message
news:MP************************@news1.eircom.net.. .
In article <44***************@yahoo.com>, cb********@yahoo.com says...
These assumptions are generally made because of familiarity with
the language. As a non-code example, consider the idea that the
faulty code is written by blackguards bent on foulling the
language. The term blackguards is not in favor these days, and for
good reason.


About as good a reason as the term niggardly, as far as I can tell.
Perhaps the words are appropriate in a post relating to fatal
assumptions.


I didn't know what he meant either. Not being racist (at least I hope not),
I went GIYF'ing. I think it might be a reference to some Dungeons and
Dragons persona or something. Unfortunately, he'd need to clarify...
Rod Pemberton
Mar 8 '06 #43

P: n/a
In article <pa****************************@areilly.bpc-users.org> Andrew Reilly <an*************@areilly.bpc-users.org> writes:
....
It's precicely this sort of tomfoolery on the part of the C standards
committee that has brought the language into such ill-repute in recent
years. It's practically unworkable now, compared to how it was in (say)
the immediately post-ANSI-fication years.


I do not understand this. The very same restriction on pointer arithmetic
was already in the very first ANSi C standard.
--
dik t. winter, cwi, kruislaan 413, 1098 sj amsterdam, nederland, +31205924131
home: bovenover 215, 1025 jn amsterdam, nederland; http://www.cwi.nl/~dik/
Mar 8 '06 #44

P: n/a
In article <pa****************************@areilly.bpc-users.org> Andrew Reilly <an*************@areilly.bpc-users.org> writes:
....
OK, I've made enough of a fool of myself already. I'll go and have that
second cup of coffee for the morning, before I start going on about having
the standard support non-2's complement integers, or machines that have no
arithmetic right shifts...


In the time of the first standard, the Cray-1 was still quite important,
and it has no arithmetic right shift. When K&R designed C there was a
large number of machines that did not use 2's complement integers.
--
dik t. winter, cwi, kruislaan 413, 1098 sj amsterdam, nederland, +31205924131
home: bovenover 215, 1025 jn amsterdam, nederland; http://www.cwi.nl/~dik/
Mar 8 '06 #45

P: n/a
Andrew Reilly <an*************@areilly.bpc-users.org> writes:
On Tue, 07 Mar 2006 23:57:38 +0000, Keith Thompson wrote:
Consider the alternative.

#define LEN 100
#define INC 5000
int arr[LEN];
int *ptr = arr;
/* A */
ptr += 2*INC;
ptr -= INC;
ptr -= INC;
/* B */
ptr -= INC;
ptr -= INC;
ptr += 2*INC;
/* C */

What you're suggesting, I think, is that ptr==arr should be true at points
A, B, and C, for any(?) values of LEN and INC. It happens to work out
that way sometimes (at least in the test program I just tried), but I can
easily imagine a system where guaranteeing this would place an undue
burden on the compiler.


That *is* what I'm suggesting. In fact, I'm suggesting that p += a; p -=
a; should leave p as it was originally for any int a and pointer p. To my
mind, and I've been using C for more than 20 years, that is the very
essence of the nature of C.

[snip]

There's (at least) one more property I forgot to mention. Given:

#define LEN 100
#define INC 5000 /* modify both of these as you like */
int arr[LEN];
int *ptr1 = arr;
int *ptr2 = ptr1 + INC;
/* D */

would you also require that, at point D, ptr2 > ptr1? (If pointer
arithmetic wraps around, this might not be the case even if adding and
subtracting as above always gets you back to the original address.)

And you think that having the standard guarantee this behavior is
worth the cost of making it much more difficult to implement C on
systems where the underlying machine addresses don't meet this
property, yes?

If so, that's a consistent point of view, but I disagree with it.

I'll also mention that none of this stuff has changed significantly
between C90 (the 1990 ISO C standard, equivalent to the original ANSI
standard of 1989) and C99 (the 1990 ISO standard).

In fact, I just checked my copy of K&R1 (published in 1978). I can't
copy-and-paste from dead trees, so there may be some typos in the
following. This is from Appendix A, the C Reference Manual, section
7.4, Additive operators:

A pointer to an object in an array and a value of any integral
type may be added. [...] The result is a pointer of the same
type as the original pointer, and which points to another object
in the same array, appropriately offset from the orginal object.

[...]

[... likewise for subtracting an integer from a pointer ...]

If two pointers to objects of the same type are subtracted, the
result is converted [...] to an int representing the number of
objects separating the pointed-to objects. This conversion will
in general give unexpected results unless the pointers point to
objects in the same array, since pointers, even to objects of the
same type, do not necessarily differ by a multiple of the
object-length.

The last quoted paragraph isn't quite as strong as what the current
standard says, since it bases the undefinedness of pointer subtraction
beyond the bounds of an object on alignment, but it covers the same
idea.

The C Reference Manual from May 1975,
<http://cm.bell-labs.com/cm/cs/who/dmr/cman.pdf>, has the same wording
about pointer subtraction, but not about pointer+integer addition.

So if you think that the requirements you advocate are "the very
essence of the nature of C", I'm afraid you're at least 28 years too
late to do anything about it.

--
Keith Thompson (The_Other_Keith) ks***@mib.org <http://www.ghoti.net/~kst>
San Diego Supercomputer Center <*> <http://users.sdsc.edu/~kst>
We must do something. This is something. Therefore, we must do this.
Mar 8 '06 #46

P: n/a
On 2006-03-07, James Dow Allen <jd*********@yahoo.com> wrote:
[...] but I'm sincerely curious whether anyone knows of an *actual*
environment where p == s will ever be false after (p = s-1; p++).
The problem is that evaluating s-1 might cause an underflow and a
trap, and then you won't even reach the comparison. You don't
necessarily have to dereference an invalid pointer to get a trap.

You might hit this behavior on any segmented architecture (e.g.,
80286, or 80386+ with segments on) and you are virtually guaranteed to
hit it on any architecture with fine-grained segmentation. comp.std.c
periodically reminisces about the old Burroughs architecture, and
it's always possible something like it might come back sometime.

You will also see this behavior in any worthwhile bounds-checking
implementation.
Many of the discussions in comp.lang.c seem like they'd be better
in a new newsgroup:
comp.lang.i'd_rather_be_a_lawyer

:-) :-)


Yes, well, that's what comp.lang.c is about...

--
- David A. Holland
(the above address works if unscrambled but isn't checked often)
Mar 8 '06 #47

P: n/a
On Wed, 08 Mar 2006 00:56:03 +0000, Robin Haigh wrote:
It's not always equivalent. The trouble starts with

char a[8];
char *p;

for ( p = a+1 ; p < a+8 ; p += 2 ) {}

intending that the loop terminates on p == a+9 (since it skips a+8). But
how do we know that a+9 > a+8 ? If the array is right at the top of some
kind of segment, the arithmetic might have wrapped round.
a+9 > a+8 because a + 9 - (a + 8) == 1, which is > 0. Doesn't matter if
the signed or unsigned pointer value wrapped around in an intermediate
term. On many machines that's how the comparison is done anyway. You're
suggesting that having the compiler ensure that a+8 doesn't wrap around
wrt a is OK, but a+9 is too hard. I don't buy it.
To support predictable pointer comparisons out of range, the compiler
would have to allocate space with a safe buffer zone. Complications are
setting in.
Only if you put them there. (The real problem is objects larger than half
the address space, where a valid pointer difference computation produces a
ptrdiff value that is out of range for a signed integer.)
Ints have the nice property that 0 is in the middle and we know how much
headroom we've got either side. So it's easy for the compiler to make
the int version work (leaving it to the programmer to take
responsibility for avoiding overflow, which is no big deal).
Unsigned ints have the nice property that (a + 1) - 1 == a for all a, even
if a + 1 == 0. Overflow is generally no big deal in any case. (Other
than the object larger than half the address space issue.)
Pointers don't have that property. The compiler can't take sole
responsibility for avoiding overflow irrespective of what the programmer
does. If the programmer wants to go out of range and is at the same
time responsible for avoiding overflow, then he has to start worrying
about whereabouts his object is and what headroom he's got.


The compiler can't necessarily avoid overflow, but it *can* arrange for
pointer comparisons to work properly.
Bah, humbug. Think I'll go back to assembly language, where pointers
do what you tell them to, and don't complain about it to their lawyers.


Can't see how assembly programmers avoid the same kind of issue. I can
see how they could ignore it. The code will work most of the time.


Seems like it will work at least as well as the usual unit-stride
algorithm and idiom.

--
Andrew

Mar 8 '06 #48

P: n/a

"Andrew Reilly" <an*************@areilly.bpc-users.org> wrote in message
news:pa****************************@areilly.bpc-users.org...
On Tue, 07 Mar 2006 13:28:37 -0500, Arthur J. O'Dwyer wrote:
K&R answers your question. If pa points to some element of an array,
then pa-1 points to the /previous element/. But what's the "previous
element" relative to the first element in the array? It doesn't exist. So we have undefined behavior.


Only because the standard says so. Didn't have to be that way. There are
plenty of logically correct algorithms that could exist that involve
pointers that point somewhere outside of a[0..N]. As long as there's no
de-referencing, no harm, no foul. (Consider the simple case of iterating
through an array at a non-unit stride, using the normal p < s + N
termination condition. The loop finishes with p > s + N and the standard
says "pow, you're dead"


and if the arithmetic happens to wrap round after s + N, you really are dead
too.

It doesn't have to be about weird architectures and traps. No
implementation can provide an unlimited range for pointer arithmetic without
some kind of overflow behaviour, such as a wrap round. Granted a wrap-round
needn't affect addition and subtraction, but it will affect comparisons.

Every allocated object comes with a limited range for pointer comparisons to
satisfy p-1<p<p+1. Not because the standard says so, but because the
implementation can't avoid it.

The kind of code you're talking about tends to make careless assumptions
about the valid range with no justification at all. Just because we've all
been doing it for years (I don't mind pleading guilty) doesn't make it
right. Such code is broken, and it doesn't need a standard to say so.

Fortunately some people have learnt a bit since the good old days, and they
work to higher standards now.

--
RSH


Mar 8 '06 #49

P: n/a
On 2006-03-08, Andrew Reilly <an*************@areilly.bpc-users.org> wrote:
if (buffer_pos + amount_needed > buffer_limit) {
... enlarge the buffer ...
}
I haven't used an AS/400 myself, either, but this is almost certainly the
sort of perfectly reasonable code that the standard has arranged to be
undefined, precicely so that it can be said that there's a C compiler for
that system.


There are lots of embedded systems with 8- and 16-bit pointers. With
the right value of buffer_pos, it wouldn't take a very large value of
amount_needed for that addition to wrap and given you an incorrect
comparison.

--
John W. Temples, III
Mar 8 '06 #50

351 Replies

This discussion thread is closed

Replies have been disabled for this discussion.