Re: [tree-ssa] Misc C++ improvements

[law at redhat dot com]

In message <1048278341 dot 2832 dot 16 dot camel at tornado>, Diego Novillo writes:
 >On Fri, 2003-03-21 at 15:17, Jason Merrill wrote:
 >
 >> Yes, all classes have a nested TYPE_DECL referring to the class itself.
 >> But my question remains: Why does the VLA code care about TYPE_DECLs?  We
 >> should only be interested in the types of local variables, specifically
 >> either ARRAY_TYPEs or RECORD_TYPEs which contain FIELD_DECLs of ARRAY_TYPE
 >> (or RECORD_TYPE, etc.).  There's no reason to look at TYPE_DECLs at all.
 >> 
 >I cannot remember the details now, but when I implemented the hack that
 >looked for VLA in array declarations I also had to look inside typedefs
 >that contained dynamic arrays.  I seem to remember this being needed for
 >gcc.c-torture/execute/20020412-1.c
20020412-1.c certainly doesn't need to be looking at the TYPE_DECLs,
nor does any other test that I can find.

Are you sure you weren't working around a bug elsewhere, particularly in
the life analysis code that went in at just about the same time?

If you recall we were failing to consider the uses that occurred in
VDEFs when computing what variables were live across block boundaries.
That caused us to mis-compile some tests, including 20020412-1.c if
I recall correctly.

Jeff