diff mbox

[RFC] combine: Handle zero_extend without subreg in change_zero_ext.

Message ID 20170105164651.GA6294@linux.vnet.ibm.com
State New
Headers show

Commit Message

Dominik Vogt Jan. 5, 2017, 4:46 p.m. UTC 
The attached patch deals with another type of zero_extend that is
not yet handled in change_zero_ext, i.e. (zero_extend
(pseudoreg)), without a "subreg" in between.  What do you think?
(Mostly untested yet.)

Ciao

Dominik ^_^  ^_^

Comments

Segher Boessenkool Jan. 11, 2017, 6:02 p.m. UTC | #1 
On Thu, Jan 05, 2017 at 05:46:51PM +0100, Dominik Vogt wrote:
> The attached patch deals with another type of zero_extend that is
> not yet handled in change_zero_ext, i.e. (zero_extend
> (pseudoreg)), without a "subreg" in between.  What do you think?
> (Mostly untested yet.)

My main question is: where is this useful?  Can you show some example
please?


Segher
Dominik Vogt Jan. 12, 2017, 10:24 a.m. UTC | #2 
On Wed, Jan 11, 2017 at 12:02:40PM -0600, Segher Boessenkool wrote:
> On Thu, Jan 05, 2017 at 05:46:51PM +0100, Dominik Vogt wrote:
> > The attached patch deals with another type of zero_extend that is
> > not yet handled in change_zero_ext, i.e. (zero_extend
> > (pseudoreg)), without a "subreg" in between.  What do you think?
> > (Mostly untested yet.)
> 
> My main question is: where is this useful?  Can you show some example
> please?

With this test program:

  int foo (int *b)
  {
    int i;
    int r;
    for (i = 0; i < 77; i++)
    {
      r |= b[0] > 0;
      r |= b[i] > 0;
    }
    return r;
  }

before combine we have

  (insn 47 45 48 7 (set (reg:SI 104)
          (zero_extend:SI (reg:QI 103)))
  (insn 48 47 49 7 (parallel [
              (set (reg/v:SI 88 [ r ])
                  (ior:SI (reg/v:SI 88 [ r ])
                      (reg:SI 104)))
              (clobber (reg:CC 33 %cc))
          ])

combine tries

  (set (reg/v:SI 88 [ r ])
      (ior:SI (zero_extend:SI (reg:QI 103))
          (reg/v:SI 88 [ r ])))

With the patch it also tries

(set (reg/v:SI 88 [ r ])
    (ior:SI (and:SI (subreg:SI (reg:QI 103) 0)
            (const_int -1 [0xffffffffffffffff]))
        (reg/v:SI 88 [ r ])))

which is just one of the standard patterns for rosbg.

Ciao

Dominik ^_^  ^_^
Segher Boessenkool Jan. 12, 2017, 5:02 p.m. UTC | #3 
Hi Dominik,

Thanks for the example.  ROSBG, what a weird instruction.

On Thu, Jan 05, 2017 at 05:46:51PM +0100, Dominik Vogt wrote:
> --- a/gcc/combine.c
> +++ b/gcc/combine.c
> @@ -11372,6 +11372,16 @@ change_zero_ext_src (subrtx_ptr_iterator *piter)
>    else if (GET_CODE (x) == ZERO_EXTEND
>  	   && SCALAR_INT_MODE_P (mode)
>  	   && REG_P (XEXP (x, 0))
> +	   && !HARD_REGISTER_P (XEXP (x, 0))
> +	   && GET_MODE_PRECISION (GET_MODE (XEXP (x, 0)))
> +	      < GET_MODE_PRECISION (mode))
> +    {
> +      /* (zero_extract (reg)) -> (and (subreg (reg) 0) (const_int)) */

s/zero_extract/zero_extend/

> 	* combine.c (change_zero_ext_src): Handle zero_extend without subreg.

"Handle zero_extend of a pseudo."?

Okay for trunk with that.  Thanks for the patch,


Segher
Dominik Vogt Jan. 12, 2017, 6:28 p.m. UTC | #4 
On Thu, Jan 12, 2017 at 11:02:36AM -0600, Segher Boessenkool wrote:
> Hi Dominik,
> 
> Thanks for the example.  ROSBG, what a weird instruction.
> 
> On Thu, Jan 05, 2017 at 05:46:51PM +0100, Dominik Vogt wrote:
> > --- a/gcc/combine.c
> > +++ b/gcc/combine.c
> > @@ -11372,6 +11372,16 @@ change_zero_ext_src (subrtx_ptr_iterator *piter)
> >    else if (GET_CODE (x) == ZERO_EXTEND
> >  	   && SCALAR_INT_MODE_P (mode)
> >  	   && REG_P (XEXP (x, 0))
> > +	   && !HARD_REGISTER_P (XEXP (x, 0))
> > +	   && GET_MODE_PRECISION (GET_MODE (XEXP (x, 0)))
> > +	      < GET_MODE_PRECISION (mode))
> > +    {
> > +      /* (zero_extract (reg)) -> (and (subreg (reg) 0) (const_int)) */
> 
> s/zero_extract/zero_extend/
> 
> > 	* combine.c (change_zero_ext_src): Handle zero_extend without subreg.
> 
> "Handle zero_extend of a pseudo."?

k

> Okay for trunk with that.

It still needs testing.  I'll do that in a while.

Ciao

Dominik ^_^  ^_^
Dominik Vogt Jan. 23, 2017, 8:38 a.m. UTC | #5 
On Thu, Jan 12, 2017 at 11:02:36AM -0600, Segher Boessenkool wrote:
> Thanks for the example.  ROSBG, what a weird instruction.

And it's annoying too because it does the same as RISBG in some
cases and can confuse "combine" to take a different branch that
turns out to be a dead end.

> On Thu, Jan 05, 2017 at 05:46:51PM +0100, Dominik Vogt wrote:
> > --- a/gcc/combine.c
> > +++ b/gcc/combine.c
> > @@ -11372,6 +11372,16 @@ change_zero_ext_src (subrtx_ptr_iterator *piter)
> >    else if (GET_CODE (x) == ZERO_EXTEND
> >  	   && SCALAR_INT_MODE_P (mode)
> >  	   && REG_P (XEXP (x, 0))
> > +	   && !HARD_REGISTER_P (XEXP (x, 0))
> > +	   && GET_MODE_PRECISION (GET_MODE (XEXP (x, 0)))
> > +	      < GET_MODE_PRECISION (mode))
> > +    {
> > +      /* (zero_extract (reg)) -> (and (subreg (reg) 0) (const_int)) */
> 
> s/zero_extract/zero_extend/
> 
> > 	* combine.c (change_zero_ext_src): Handle zero_extend without subreg.
> 
> "Handle zero_extend of a pseudo."?
> 
> Okay for trunk with that.  Thanks for the patch,

The patch forgot to set "size" and generated invalid Rtl because
of that.  Fixed and tested, but I'd prefer to repost my growing
stack of patches when gcc-7 is released.  There are too many
changes in "combine" and "simplify" that may depend on each other.

Ciao

Dominik ^_^  ^_^
diff mbox

Patch

diff --git a/gcc/combine.c b/gcc/combine.c
index e77f203..b5131b8 100644
--- a/gcc/combine.c
+++ b/gcc/combine.c
@@ -11372,6 +11372,16 @@  change_zero_ext_src (subrtx_ptr_iterator *piter)
   else if (GET_CODE (x) == ZERO_EXTEND
 	   && SCALAR_INT_MODE_P (mode)
 	   && REG_P (XEXP (x, 0))
+	   && !HARD_REGISTER_P (XEXP (x, 0))
+	   && GET_MODE_PRECISION (GET_MODE (XEXP (x, 0)))
+	      < GET_MODE_PRECISION (mode))
+    {
+      /* (zero_extract (reg)) -> (and (subreg (reg) 0) (const_int)) */
+      x = gen_lowpart_SUBREG (mode, XEXP (x, 0));
+    }
+  else if (GET_CODE (x) == ZERO_EXTEND
+	   && SCALAR_INT_MODE_P (mode)
+	   && REG_P (XEXP (x, 0))
 	   && HARD_REGISTER_P (XEXP (x, 0))
 	   && can_change_dest_mode (XEXP (x, 0), 0, mode))
     {