RE: [PATCH 1/2] compiler.h: Introduce ptr_eq() to preserve address dependency

[David Laight <David.Laight@ACULAB.COM>]

From: 'Alan Stern'
> Sent: 02 October 2024 15:15
> 
> On Wed, Oct 02, 2024 at 08:13:15AM +0000, David Laight wrote:
> > From: 'Alan Stern'
> > > Sent: 01 October 2024 23:57
> > >
> > > On Tue, Oct 01, 2024 at 05:11:05PM +0000, David Laight wrote:
> > > > From: Alan Stern
> > > > > Sent: 30 September 2024 19:53
> > > > >
> > > > > On Mon, Sep 30, 2024 at 07:05:06PM +0200, Jonas Oberhauser wrote:
> > > > > >
> > > > > >
> > > > > > Am 9/30/2024 um 6:43 PM schrieb Alan Stern:
> > > > > > > On Mon, Sep 30, 2024 at 01:26:53PM +0200, Jonas Oberhauser wrote:
> > > > > > > >
> > > > > > > >
> > > > > > > > Am 9/28/2024 um 4:49 PM schrieb Alan Stern:
> > > > > > > >
> > > > > > > > I should also point out that it is not enough to prevent the compiler from
> > > > > > > > using @a instead of @b.
> > > > > > > >
> > > > > > > > It must also be prevented from assigning @b=@a, which it is often allowed to
> > > > > > > > do after finding @a==@b.
> > > > > > >
> > > > > > > Wouldn't that be a bug?
> > > > > >
> > > > > > That's why I said that it is often allowed to do it. In your case it
> > > > > > wouldn't, but it is often possible when a and b are non-atomic &
> > > > > > non-volatile (and haven't escaped, and I believe sometimes even then).
> > > > > >
> > > > > > It happens for example here with GCC 14.1.0 -O3:
> > > > > >
> > > > > > int fct_hide(void)
> > > > > > {
> > > > > >     int *a, *b;
> > > > > >
> > > > > >     do {
> > > > > >         a = READ_ONCE(p);
> > > > > >         asm volatile ("" : : : "memory");
> > > > > >         b = READ_ONCE(p);
> > > > > >     } while (a != b);
> > > > > >     OPTIMIZER_HIDE_VAR(b);
> > > > > >     return *b;
> > > > > > }
> > > > > >
> > > > > >
> > > > > >
> > > > > >         ldr     r1, [r2]
> > > > > >         ldr     r3, [r2]
> > > > > >         cmp     r1, r3
> > > > > >         bne     .L6
> > > > > >         mov     r3, r1   // nay...
> > > > >
> > > > > A totally unnecessary instruction, which accomplishes nothing other than
> > > > > to waste time, space, and energy.  But nonetheless, allowed -- I agree.
> > > > >
> > > > > The people in charge of GCC's optimizer might like to hear about this,
> > > > > if they're not already aware of it...
> > > > >
> > > > > >         ldr     r0, [r3] // yay!
> > > > > >         bx      lr
> > > > >
> > > > > One could argue that in this example the compiler _has_ used *a instead
> > > > > of *b.  However, such an argument would have more force if we had
> > > > > described what we are talking about more precisely.
> > > >
> > > > The 'mov r3, r1' has nothing to do with 'a'.
> > >
> > > What do you mean by that?  At this point in the program, a is the
> > > variable whose value is stored in r1 and b is the variable whose value
> > > is stored in r3.  "mov r3, r1" copies the value from r1 into r3 and is
> > > therefore equivalent to executing "b = a".  (That is why I said one
> > > could argue that the "return *b" statement uses the value of *a.)  Thus
> > > it very much does have something to do with "a".
> >
> > After the cmp and bne r1 and r3 have the same value.
> > The compiler tracks that and will use either register later.
> > That can never matter.
> 
> The whole point of this thread is that sometimes it _does_ matter.  Not
> on x86, but on weakly ordered architectures where using the wrong
> register will bypass a dependency and allow the CPU to speculatively
> load values earlier than the programmer wants it to.
> 
> > Remember the compiler tracks values (in pseudo/internal registers)
> > not variables.
> >
> > > > It is a more general problem that OPTIMISER_HIDE_VAR() pretty much
> > > > always ends up allocating a different internal 'register' for the
> > > > output and then allocating a separate physical rehgister.
> > >
> > > What output are you referring to?  Does OPTIMISER_HIDE_VAR() have an
> > > output?  If it does, the source program above ignores it, discarding any
> > > returned value.
> >
> > Look up OPTIMISER_HIDE_VAR(x) it basically x = f(x) where f() is
> > the identity operation:
> > 	asm ("" : "+r"(x))
> > I'll bet that gcc allocates a separate internal/pseudo register
> > for the result so wants to do y = f(x).
> > Probably generating y = x; y = f(y);
> > (The 'mov' might be after the asm, but I think that would get
> > optimised away - the listing file might help.)
> >
> > So here the compiler has just decided to reuse the register that
> > held the other of a/b for the extra temporary.
> 
> I think you've got this backward.  As mentioned above, a is originally
> in r1 and b is in r3.  The source says OPTIMIZER_HIDE_VAR(b), so you're
> saying that gcc should be copying r3 into a separate internal/pseudo
> register.  But instead it's copying r1.

I think I know what you are trying to do, and you just fail.
Whether something can work is another matter, but that code
can't ever work.

Inside if (a == b) the compiler will always use the same register
for references to a and b - because it knows they have the same value.

Possibly something like:
	c = b;
	OPTIMISER_HIDE_VAR(c);
	if (a == c) {
		*b
will ensure that there isn't a speculative load from *a.
You'll get at least one register-register move - but they are safe.
Otherwise you'll need to put the condition inside an asm block.

	David

> 
> Alan

-
Registered Address Lakeside, Bramley Road, Mount Farm, Milton Keynes, MK1 1PT, UK
Registration No: 1397386 (Wales)