[Bug tree-optimization/77485] Missed dead store elimination of aggregate store followed by partial stores

[rguenther at suse dot de]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=77485

--- Comment #8 from rguenther at suse dot de <rguenther at suse dot de> ---
On Wed, 14 Dec 2016, law at redhat dot com wrote:

> https://gcc.gnu.org/bugzilla/show_bug.cgi?id=77485
> 
> --- Comment #7 from Jeffrey A. Law <law at redhat dot com> ---
> There was a bit of cruft and a missed ADDR_EXPR in that last fragment of gimple
> code.
> 
> ;;   basic block 2, loop depth 0, count 0, freq 0, maybe hot
> ;;    prev block 0, next block 1, flags: (NEW, REACHABLE, VISITED)
> ;;    pred:       ENTRY (FALLTHRU,EXECUTABLE)
>   __builtin_memset (&MEM[(struct FixBuf *)&<retval> + 31B], 0, 169);
>   <retval>.buf[0] = 48;
>   <retval>.buf[1] = 32;
> [ ... ]
>   return <retval>;

It also shows another (awkward) detail of the transform -- a previously
non-TREE_ADDRESSABLE <retval> becomes TREE_ADDRESSABLE.  This 
generally invalidates points-to (though in this particular case should
be harmless).  It also possibly pessimizes alias analysis because
pointers pointing to unknown stuff now also alias <retval> while
previously not (non-TREE_ADDRESSABLE decls are not aliased).

This is why we generally prefer

 decl = {};

over a memset.  If you just chop off elements from the head/tail
you can keep the {} if you do

 MEM <char[of-correct-size]> [&decl, off with original alias type]
   = {}; (CONSTRUCTOR of same char[] type)