Re: [Firebird-devel] BLK* to void** implicit conversion problem

[John B. <bel...@cs...>]

On Thursday, June 14, 2001, at 03:09 PM, Ann W. Harrison wrote:

> At 02:27 PM 6/14/2001 -0700, John Bellardo wrote:
>> [.....]
>> Gpre produces a number of calls to internal functions (such as 
>> isc_compile_request) that are expecting void**.  For the time being 
>> I've been changing the BLK* declaration to void**, and commenting it 
>> to fix later.  What is the correct fix?  Do the isc_compile_request, 
>> et. al functions expect a BLK* in void** clothing, or do the really 
>> operate on a void**.  I could define a c++ type operator that converts 
>> a blk->void*, but that won't help with the **'s.
>> Any thoughts?  Thanks.
>
>
> For starters, isc_compile_request is not an internal function, but part
> of the previously documented API.

My mistake.

> When gpre is using that level of interface,
> it should be generating request handle declarations like this:
>
>                 static isc_req_handle
>                         isc_7 = 0;              /* request handle */
>
> I believe that isc_req_handle et al. resolve down to void **, though I 
> would
> prefer to see the routines declared with the actual handle type-names 
> for
> the sake of documentation.  They were changed to void ** by the previous
> administration.

Sounds good, but I'm using the -g switch (see below) so I don't think it 
applies...

>
> However, for dfw.e and other .e files within the engine, gpre should be 
> generating
> actual internal calls to routines like CMP_compile2 request
>
>         request = (BLK) CMP_compile2 (tdbb, jrd_356, TRUE);
>
> If the code you're seeing in DFW is not like this, you should use the -g
> switch.

And that is what I am seeing.  Here is a the first part of the 
DFW_update_index function from dfw.epp:

         BLK request;

         tdbb = GET_THREAD_DATA;
         dbb = tdbb->tdbb_database;

         request = (BLK) CMP_find_request(tdbb, irq_m_index, 
IRQ_REQUESTS);

         FOR(REQUEST_HANDLE request)
                 IDX IN RDB$INDICES WITH IDX.RDB$INDEX_NAME EQ 
work->dfw_name
      .........


And the resulting code (after gpre) from dfw.cpp:

         BLK request;

         tdbb = GET_THREAD_DATA;
         dbb = tdbb->tdbb_database;

         request = (BLK) CMP_find_request(tdbb, irq_m_index, 
IRQ_REQUESTS);

         /*FOR(REQUEST_HANDLE request)
                 IDX IN RDB$INDICES WITH IDX.RDB$INDEX_NAME EQ 
work->dfw_name*/
         {
         {
         if (!DB)
            isc_attach_database ((long*) 0L, 0, "ODS.RDB", &DB, 0, 
(char*) 0);
         if (DB && !gds__trans)
            isc_start_transaction ((long*) 0L, &gds__trans, (short) 1, 
&DB, (shor
t) 0, (char*) 0);
         }
         if (!request)
            isc_compile_request ((long*) 0L, &DB, &request, (short) 
sizeof (isc_3
56), (char ISC_FAR *) isc_356);

    ....................

I've included enough code to see the first offender; the call to 
isc_compile_request third parameter.  The request variable is a BLK (ie, 
blk*), so &request is a blk**.  isc_compile_request expects a void**.  I 
looked in a copy of Stroustrup and there is an implicit conversion from 
xxx* to void*, but that doesn't apply because we are using blk** to 
void**.  Since the call generating the error is created from gpre, 
should gpre always include a cast to void** when it emits the calls?  
That would be trusting the user to do the right thing.  Otherwise I need 
to continue changing the declaration of "request" to void* from BLK.

>
> There's more, of course - the blk structure is a significant part of the
> internal memory management and as such, quite dear to me.  What we 
> should
> do is ditch all the memory management and use destructors and use counts
> and that sort of thing - but that's a project and a half.  Until then, 
> the
> blk structure is of some considerable importance.

One can always dream (or make a wish list).

-John