Re: [PATCH] arch: tile: kernel: kgdb.c: Use memcpy() instead of pointer copy one by one

[Jeff Epler <jepler@unpythonic.net>]

On Wed, Nov 12, 2014 at 11:43:08PM +0800, Chen Gang wrote:
> >                                                       (I assume the
> > compiler could do things like replace an intended load from memory with
> > a constant load or even no load at all)
> > 
> 
> Excuse me, my English is not quite well, I can not understand what you
> said above. (If necessary, please help provide more details for it).

I am concerned that writing regs[TREG_TP] is "undefined behavior"
according to the C standard.

This expression is equivalent to *(regs + TREG_TP).  The expression
(regs + TREG_TP) does not result in a pointer to any element of regs[],
so dereferencing it is undefined behavior.  (Source: C99 draft standard
WG14/N1256, annex J.2, "[The behavior is undefined if t]he operand of
the unary * operator has an invalid value")

That is why the compiler showed the original diagnostic, but the same
logic that made the loop's behavior undefined also makes the expression
regs[TREG_TP] undefined whereever it appears.

None of this is a specific problem with your proposed patch.  Rather, it
is a suggestion that the whole structure's design needs to be revisited
in light of compilers beginning to notice that regs[TREG_TP] is
undefined behavior and change their generated code as a result.

Unfortunately it looks like this header is also a part of the userspace
API, so it can't simply be changed just in case all in-kernel uses are
changed.

Jeff