Re: [PATCH] PR c/8439

[Mark Mitchell <mark@codesourcery.com>]

> The proposed fix is to teach expand_increment not to emit insn when the
> increment is null. Bootstrapped/regtested (c,c++,objc,f77 mainline) on
> i586-redhat-linux-gnu.

This fix isn't correct; if the thing being incremented is volatile, we
should still access it.

The real problem is that validate_replace_rtx_1 is performing
simplifications; when we replace the virtual argument pointer with
the real one it tries to get rid of the plus in the following:

(set (mem/f:SI (reg/f:SI 53 virtual-incoming-args) [0 p+0 S4 A32])
  (plus:SI (mem/f:SI (reg/f:SI 53 virtual-incoming-args) [0 p+0 S4 A32])
    (const_int 0 [0x0])))

which makes the expression unrecognizable.

There's even a comment:

      /* If we have a PLUS whose second operand is now a CONST_INT, use
         plus_constant to try to simplify it.
         ??? We may want later to remove this, once simplification is
         separated from this function.  */

We should really be simplifying only if *both* arguments are now
CONST_INTs.  I'll try that for mainline.

I thought about trying to fix this as you suggested, or even earlier
in the production of trees, but the volatile thing makes it a bit
tricky, and the bottom line is that validate_replace_rtx_1 is broken.

So, we'll punt this one to another day; I can't see messing with anything
in the generic middle end code to fix a bug involving the GNU zero-sized
structure extension.  (Insert rant about extensions here.)

-- 
Mark Mitchell                mark@codesourcery.com
CodeSourcery, LLC            http://www.codesourcery.com