Re: Named Pipe unformatted I/O (PR30162)

[Jerry DeLisle <jvdelisle at verizon dot net>]

Jerry DeLisle wrote:
> Hi all,
>
> In looking at the -fdump-tree-original for the test cases we have and 
> the use of allocated example given by Thomas in this thread, you will 
> see that the frontend currently generates appropriate calls to the 
> various existing transfer functions as needed.
>
> If we modify this to call, for example, transfer_integer_init instead of 
> transfer_integer, we can use the existing front-end fairly intact to 
> generate calls to new init functions that will build the output list as 
> a linked list with all the size information needed (or computable such 
> as with array descriptors) and the related data ready for transfer.
>
> Embedded function calls use temporaries to hold the results. We will 
> have to move the location of memory free calls related to those until 
> after the transfer is complete.
>
> After the complete output list is created, inside st_write_done, we 
> modify finalize_transfer to traverse the linked list and call the 
> existing transfer functions in order, with all the pertinent data ready 
> to go.
>
> This seems very doable.  I think performance will hold OK and we may 
> even spot some optimizations that we could do.
>
> Jerry
BTW, I am presenting my thoughts as I go here.  This is conceptual design phase 
so I am not locked into any particular approach yet. Just getting ideas and 
issues captured.

Jerry