[PATCH] Fix type correctness of fold_indirect_ref

[Mark Mitchell]

Richard Guenther wrote:

> Well, I'm not very familiar with side-effects the following can cause,
> but, the following patch moves the optimization to the C++ frontend,
> thereby fixing the temp1.C regression.
> 
> If this is generally ok I'll throw it at bootstrap & regtesting over night.

I do not believe this is safe.

It might be possible to do that if we kept track of lvalues and rvalues 
more carefully in the C++ front end, which we would like to do, but that 
is not done at present.  Until it is, the change you're making might 
result in replacing "*(const int*)&i" (where "i" is an "int") with "i", 
an expression of type "int", which is wrong.  For example, the original 
expression cannot be trivially bound to an "int &", but "i" can be.

Furthermore, the C++ front end probably shouldn't be doing this kind of 
optimization anyhow; we really want it to be faithful to the source 
representation to the extent reasonable.  (See extensive previous 
discussions on this point.)  It may be that the C++ gimplifier should be 
doing this, but not the front end proper.

-- 
Mark Mitchell
CodeSourcery, LLC
mark@codesourcery.com
(916) 791-8304