[PR c/84293] Unexpected strict-alias warning

[Nathan Sidwell]

This bug comes from python sources.  It has some type-punning macros to 
switch between representation, but these can cause unexpected 
strict-alias warnings, even though the punning's context is a system 
header file.  (I guess that without macro expansion location information 
we'd've thought the context was user code.)

As shown in the bug report, if the pun is spelt '(struct X *)&obj' we do 
not warn, but if it is '(X_t *)&obj' we do.  We also don't warn in 
either case if we go via -save-temps.  The patch changes things to not 
warn in either spelling.

The problem is that strict_aliasing_warning uses input_location to 
identifier the current code point.  But that's not necessarily the 
location of the above pun -- it can be the point where we use the pun.

We cannot just use the incoming EXPR's location, because we may well 
have peeled some pieces off it, and we end up with the '&obj' fragment, 
which in this case is located in the user source.  It also turns out 
that strict_aliasing_warning no longer uses the incoming 'otype' 
argument, immediately doing:
    STRIP_NOPS (expr);
    otype = TREE_TYPE (expr);

So we may as well drop that arg, substituting a location_t from the caller.

That's what this patch does.  The callers in the C & C++ FEs pass the 
EXPR_LOCATION of the pointer being dereferenced, or converted, as 
appropriate.

Joseph, are the C bits ok?

nathan

-- 
Nathan Sidwell
-------------- next part --------------
A non-text attachment was scrubbed...
Name: 84293.diff
Type: text/x-patch
Size: 7992 bytes
Desc: not available
URL: <http://gcc.gnu.org/pipermail/gcc-patches/attachments/20180209/92aaafc3/attachment.bin>