On 1/5/21 10:58 AM, Adhemerval Zanella wrote:
> The idea of removing the alloca allows a slight better code generation,
> simplifies the boilerplate code to avoid the unbounded alloca usage,
> and it plays better with security compiler mitigation tools (such as the
> ones for stack clash).

Instead of complicating dynarray by adding char_array stuff, I suggest 
adding any primitives just to glob for now, as it's not clear that 
they're generally useful.

I do see a problem with the proposed patch set, in that it creates 
several different dynarrays when glob really needs only one or two. The 
existing code is full of memory-allocation gotchas and would be 
simplified if it treated the memory it allocates as a first-class part 
of the problem rather than as some sort of cranky subsidiary that needs 
to be babied and its diaper changed at random intervals.

Attached is a draft set of patches against Gnulib commit 
6a00fdb4bb105697aa27ba97ef7ec33287790ad3 which gives a hint about the 
sort of thing that I mean here. This patch set is not complete (it does 
only the equivalent of the first four of the patches you proposed) and 
it's not exactly the form that I want, so I haven't installed it into 
Gnulib. However, I hope it shows the sort of thing I have in mind. So 
far, it's needed only one scratch buffer.

In this patch set, glob.c continues to use scratch_buffer.h because 
scratch buffers are good enough for all the changes needed so far. I 
suppose dynarrays will be helpful for later patches and that we can 
switch to them as needed, but I wanted to focus on the actual problem 
first rather than worrying about scratch buffers vs dynarrays.