is this a valid approach to aliasing?

is this a valid approach to aliasing?

[Robert William Fuller]

After twenty years of programming in C, I thought I understood the 
language pretty well until some new code I was writing started emitting 
"type-punned pointer" aliasing warnings.  So I read up on the problem so 
I would understand it.  I understand that is necessary to restrict 
aliasing in order to optimize/load stores.

My question is if my solution to the problem is valid.  Basically, I had 
some functionality I wanted to abstract which became the following function:

int reallocItems(void **items, int itemSize, int *itemAlloc, int 
numItems, int newItems, int itemHint)
{
     void *rcAlloc;
     int emptyItems, allocItems;

     // for idempotence
     allocItems = *itemAlloc;

     // will the new entries fit?
     //
     emptyItems = allocItems - numItems;
     if (newItems > emptyItems) {

         // allocate the number of entries needed or the hint, whichever 
is more
         //
         allocItems = numItems + newItems;
         if (itemHint > allocItems) {
             allocItems = itemHint;
         }

         rcAlloc = (void *) realloc(*items, allocItems * itemSize);
         if (rcAlloc) {

             *items = rcAlloc;
             *itemAlloc = allocItems;

         } else {
             return -1;
         }
     }

     return 0;
}

Naturally, when I started trying to call this function, I started 
getting the type-punned pointer aliasing warnings.  The code in question 
looked roughly like this:

static opt_param_t *shrtOpts, *longOpts;

rc = reallocItems((void **) &shrtOpts, sizeof(opt_param_t), 
&shrtOptAlloc, numShrtOpts, newShrtOpts, shrtOptHint);

Obviously, this breaks the aliasing rules.  I read that I could work 
around this by casting through a union.  I settled on this approach, but 
I'm not sure if it is valid, or if I'm merely masking the problem:

typedef union _opt_param_alias_t {

     opt_param_t *o;
     void *v;

} opt_param_alias_t;

rc = reallocItems(&((opt_param_alias_t *) &shrtOpts)->v, 
sizeof(opt_param_t), &shrtOptAlloc, numShrtOpts, newShrtOpts, shrtOptHint)

Will casting through this union force the compiler to reload the address 
of shrtOpts when I try to reference it a few lines later?  Is this 
solution correct?  Is there a better approach that does NOT involve 
changing reallocItems to return a void pointer?

I tried several different approaches, and I'm trying to find an approach 
that requires little extra code and does not ugly things up too much.

Please advise.  Thank you.

Regards,

Rob

Re: is this a valid approach to aliasing?

[Andrew Haley]

Robert William Fuller wrote:
>
> Naturally, when I started trying to call this function, I started 
> getting the type-punned pointer aliasing warnings.  The code in question 
> looked roughly like this:
> 
> static opt_param_t *shrtOpts, *longOpts;
> 
> rc = reallocItems((void **) &shrtOpts, sizeof(opt_param_t), 
> &shrtOptAlloc, numShrtOpts, newShrtOpts, shrtOptHint);
> 
> Obviously, this breaks the aliasing rules.  I read that I could work 
> around this by casting through a union.  I settled on this approach, but 
> I'm not sure if it is valid, or if I'm merely masking the problem:
> 
> typedef union _opt_param_alias_t {
> 
>     opt_param_t *o;
>     void *v;
> 
> } opt_param_alias_t;
> 
> rc = reallocItems(&((opt_param_alias_t *) &shrtOpts)->v, 
> sizeof(opt_param_t), &shrtOptAlloc, numShrtOpts, newShrtOpts, shrtOptHint)

It really isn't necessary to do this.  The rule is that you mustn't use
a pointer cast to create an lvalue of incompatible type.  So, do this:

static opt_param_t *shrtOpts, *longOpts;

void *p1;

rc = reallocItems(&p1, sizeof(opt_param_t),
    &shrtOptAlloc, numShrtOpts, newShrtOpts, shrtOptHint);
shrtOpts = p1;

Andrew.

Re: is this a valid approach to aliasing?

[Robert William Fuller]

Andrew Haley wrote:
> Robert William Fuller wrote:
>>
>> Naturally, when I started trying to call this function, I started 
>> getting the type-punned pointer aliasing warnings.  The code in 
>> question looked roughly like this:
>>
>> static opt_param_t *shrtOpts, *longOpts;
>>
>> rc = reallocItems((void **) &shrtOpts, sizeof(opt_param_t), 
>> &shrtOptAlloc, numShrtOpts, newShrtOpts, shrtOptHint);
>>
>> Obviously, this breaks the aliasing rules.  I read that I could work 
>> around this by casting through a union.  I settled on this approach, 
>> but I'm not sure if it is valid, or if I'm merely masking the problem:
>>
>> typedef union _opt_param_alias_t {
>>
>>     opt_param_t *o;
>>     void *v;
>>
>> } opt_param_alias_t;
>>
>> rc = reallocItems(&((opt_param_alias_t *) &shrtOpts)->v, 
>> sizeof(opt_param_t), &shrtOptAlloc, numShrtOpts, newShrtOpts, 
>> shrtOptHint)
> 
> It really isn't necessary to do this.  The rule is that you mustn't use
> a pointer cast to create an lvalue of incompatible type.  So, do this:
> 
> static opt_param_t *shrtOpts, *longOpts;
> 
> void *p1;
> 
> rc = reallocItems(&p1, sizeof(opt_param_t),
>    &shrtOptAlloc, numShrtOpts, newShrtOpts, shrtOptHint);
> shrtOpts = p1;

I suspected that might work.  I asked so I would not have to look at 
gcc's output.  I have not looked at (dis)assembly since I worked on 
Windows NT in its early days when things were even more closed than they 
are now.  I am now a totally reformed free software user and programmer. 
  (Ask yourself if there is a causal relation there.)

I want to make sure I am very clear on this.  Are you saying the 
following is okay?

p1 = shrtOpts;
rc = reallocItems(&p1, ...)
shrtOpts = p1;

Thank you.

Regards,

Rob

Re: is this a valid approach to aliasing?

[Andrew Haley]

Robert William Fuller wrote:
> Andrew Haley wrote:
>> Robert William Fuller wrote:
>>>
>>> Naturally, when I started trying to call this function, I started 
>>> getting the type-punned pointer aliasing warnings.  The code in 
>>> question looked roughly like this:
>>>
>>> static opt_param_t *shrtOpts, *longOpts;
>>>
>>> rc = reallocItems((void **) &shrtOpts, sizeof(opt_param_t), 
>>> &shrtOptAlloc, numShrtOpts, newShrtOpts, shrtOptHint);
>>>
>>> Obviously, this breaks the aliasing rules.  I read that I could work 
>>> around this by casting through a union.  I settled on this approach, 
>>> but I'm not sure if it is valid, or if I'm merely masking the problem:
>>>
>>> typedef union _opt_param_alias_t {
>>>
>>>     opt_param_t *o;
>>>     void *v;
>>>
>>> } opt_param_alias_t;
>>>
>>> rc = reallocItems(&((opt_param_alias_t *) &shrtOpts)->v, 
>>> sizeof(opt_param_t), &shrtOptAlloc, numShrtOpts, newShrtOpts, 
>>> shrtOptHint)
>>
>> It really isn't necessary to do this.  The rule is that you mustn't use
>> a pointer cast to create an lvalue of incompatible type.  So, do this:
>>
>> static opt_param_t *shrtOpts, *longOpts;
>>
>> void *p1;
>>
>> rc = reallocItems(&p1, sizeof(opt_param_t),
>>    &shrtOptAlloc, numShrtOpts, newShrtOpts, shrtOptHint);
>> shrtOpts = p1;
> 
> I suspected that might work.  I asked so I would not have to look at 
> gcc's output.  I have not looked at (dis)assembly since I worked on 
> Windows NT in its early days when things were even more closed than they 
> are now.  I am now a totally reformed free software user and programmer. 
>  (Ask yourself if there is a causal relation there.)
> 
> I want to make sure I am very clear on this.  Are you saying the 
> following is okay?
> 
> p1 = shrtOpts;
> rc = reallocItems(&p1, ...)
> shrtOpts = p1;

Yes.  But you shouldn't believe me!  Read the C Standard.
The section you need to understand is 6.3.2.3.

http://www.open-std.org/JTC1/SC22/WG14/www/docs/n1256.pdf

Andrew.


http://www.open-std.org/JTC1/SC22/WG14/www/docs/n1256.pdf

Re: is this a valid approach to aliasing?

[Robert William Fuller]

John Love-Jensen wrote:
> Hi Robert,
> 
> Can you just get rid of the void** for the items parameter (see below)?
> 
> --Eljay
> 
> #include <stdlib.h>
> 
> int reallocItems(
>   void* items,
>   int itemSize,
>   int* itemAlloc,
>   int numItems,
>   int newItems,
>   int itemHint)
> {
>   void* rcAlloc;
>   int emptyItems;
>   int allocItems;
> 
>   // for idempotence
>   allocItems = *itemAlloc;
> 
>   // will the new entries fit?
>   emptyItems = allocItems - numItems;
> 
>   if (newItems > emptyItems)
>   {
>     // allocate the number of entries needed or the hint, whichever is more
>     allocItems = numItems + newItems;
> 
>     if (itemHint > allocItems)
>     {
>       allocItems = itemHint;
>     }
> 
>     rcAlloc = realloc(items, allocItems * itemSize);
> 
>     if (rcAlloc)
>     {
>       *(void**)items = rcAlloc;
>       *itemAlloc = allocItems;
>     }
>     else
>     {
>       return -1;
>     }
>   }
> 
>   return 0;
> }

Would this circumvent the type system?  Is there still an alias problem 
with your suggestion, albeit without a warning?  I am not sure.

Rob

Re: is this a valid approach to aliasing?

[Andrew Haley]

Robert William Fuller wrote:
> John Love-Jensen wrote:
>> Hi Robert,
>>
>> Can you just get rid of the void** for the items parameter (see below)?
>>
>> --Eljay
>>
>> #include <stdlib.h>
>>
>> int reallocItems(
>>   void* items,
>>   int itemSize,
>>   int* itemAlloc,
>>   int numItems,
>>   int newItems,
>>   int itemHint)
>> {
>>   void* rcAlloc;
>>   int emptyItems;
>>   int allocItems;
>>
>>   // for idempotence
>>   allocItems = *itemAlloc;
>>
>>   // will the new entries fit?
>>   emptyItems = allocItems - numItems;
>>
>>   if (newItems > emptyItems)
>>   {
>>     // allocate the number of entries needed or the hint, whichever is 
>> more
>>     allocItems = numItems + newItems;
>>
>>     if (itemHint > allocItems)
>>     {
>>       allocItems = itemHint;
>>     }
>>
>>     rcAlloc = realloc(items, allocItems * itemSize);
>>
>>     if (rcAlloc)
>>     {
>>       *(void**)items = rcAlloc;

This code is probably wrong.  Make item a void** and get rid of the cast.

Andrew.