[Patch match.pd] Fold (A / (1 << B)) to (A >> B)

[Richard Biener]

On Fri, 16 Jun 2017, James Greenhalgh wrote:

> 
> On Mon, Jun 12, 2017 at 03:56:25PM +0200, Richard Biener wrote:
> > On Mon, 12 Jun 2017, James Greenhalgh wrote:
> >
> > >
> > > Hi,
> > >
> > > As subject, for the testcase in the patch:
> > >
> > >   unsigned long
> > >   f2 (unsigned long a, int b)
> > >   {
> > >     unsigned long x = 1UL << b;
> > >     return a / x;
> > >   }
> > >
> > > We currently generate:
> > >
> > >   f2:
> > > 	mov	x2, 1
> > > 	lsl	x1, x2, x1
> > > 	udiv	x0, x0, x1
> > > 	ret
> > >
> > > Which could instead be transformed to:
> > >
> > >   f2:
> > > 	lsr	x0, x0, x1
> > > 	ret
> > >
> > > OK?
> >
> > +   We can't do the same for signed A, as it might be negative, which
> > would
> > +   introduce undefined behaviour.  */
> >
> > huh, AFAIR it is _left_ shift of negative values that invokes
> > undefined behavior.
> 
> You're right this is not a clear comment. The problem is not undefined
> behaviour, so that text needs to go, but rounding towards/away from zero
> for signed negative values. Division will round towards zero, arithmetic
> right shift away from zero. For example in:
> 
>     -1 / (1 << 1)   !=    -1 >> 1
>   = -1 / 2
>   = 0                     = -1
> 
> I've rewritten the comment to make it clear this is why we can only make
> this optimisation for unsigned values.

Ah, of course.  You could use

 if ((TYPE_UNSIGNED (type)
      || tree_expr_nonnegative_p (@0))

here as improvement.

> See, for example, gcc.c-torture/execute/pr34070-2.c
> 
> > Note that as you are accepting vectors you need to make sure the
> > target actually supports arithmetic right shift of vectors
> > (you only know it supports left shift and division -- so it might
> > be sort-of-superfluous to check in case there is no arch that supports
> > those but not the other).
> 
> I've added a check for that using optabs, is that the right way to do this?

+      && (!VECTOR_TYPE_P (type)
+          || optab_for_tree_code (RSHIFT_EXPR, type, optab_vector)
+          || optab_for_tree_code (RSHIFT_EXPR, type, optab_scalar)))

is not enough -- you need sth like

 optab ot = optab_for_tree_code (RSHIFT_EXPR, type, optab_vector);
 if (ot != unknown_optab
     && optab_handler (ot, TYPE_MODE (type)) != CODE_FOR_nothing)
   .. ok! ...

ideally we'd have a helper for this in optab-tree.[ch], 
tree-vect-patterns.c could also make use of that.

Thanks,
Richard.


> Bootstrapped and tested on aarch64-none-linux-gnu with no issues.
> 
> OK?
> 
> Thanks,
> James
> 
> ---
> gcc/
> 
> 2017-06-13  James Greenhalgh  <james.greenhalgh@arm.com>
> 
> 	* match.pd (A / (1 << B) -> A >> B): New.
> 	* generic-match-head.c: Include optabs-tree.h.
> 	* gimple-match-head.c: Likewise.
> 
> gcc/testsuite/
> 
> 2017-06-13  James Greenhalgh  <james.greenhalgh@arm.com>
> 
> 	* gcc.dg/tree-ssa/forwprop-37.c: New.
> 
> 

-- 
Richard Biener <rguenther@suse.de>
SUSE LINUX GmbH, GF: Felix Imendoerffer, Jane Smithard, Graham Norton, HRB 21284 (AG Nuernberg)