[PATCH] gfortran: Improve translation of POPPAR intrinsic

[PATCH] gfortran: Improve translation of POPPAR intrinsic

[Roger Sayle]

[-- Attachment #1: Type: text/plain, Size: 1251 bytes --]

 

The following patch to gfortran's trans-instrinsic.c tweaks the generic that
is produced

for popcnt on integer(kind=16).  Currently, the double word popcnt is
implemented as

parityll(hipart(x))^parityll(lopart(x)), but with this patch this is now
translated as

parityll(hipart(x)^lopart(x)).  This will be just an aesthetic change once
my tree-level

parity optimization patch of 12th June is reviewed and accepted, but
generating the

more efficient form initially, avoids a tiny bit of garbage collection when
the middle-end

cleans this up into its preferred form.    The semantics/correctness of this

change are tested by the run-time tests in gfortran.dg/popcnt_poppar_2.F90

 

This patch has been tested with "make bootstrap" and "make -k check" on

x86_64-pc-linux-gnu with no regressions.  If approved, I'd very much

appreciate it if the (gfortran) reviewer could commit this change for me.

 

2020-06-14  Roger Sayle  <roger@bextmovesoftware.com>

 

        * trans-intrinsic.c (gfc_conv_intrinsic_popcnt_poppar): Translate

        poppar(kind=16) as parityll(hipart(x)^lopart(x)) instead of

        parityll(hipart(x))^parityll(lopart(x)).

 

 

Thanks in advance,

Roger

--

Roger Sayle

NextMove Software

Cambridge, UK

 


[-- Attachment #2: patchb.txt --]
[-- Type: text/plain, Size: 2518 bytes --]

diff --git a/gcc/fortran/trans-intrinsic.c b/gcc/fortran/trans-intrinsic.c
index fd88099..363874e 100644
--- a/gcc/fortran/trans-intrinsic.c
+++ b/gcc/fortran/trans-intrinsic.c
@@ -7153,35 +7153,39 @@ gfc_conv_intrinsic_popcnt_poppar (gfc_se * se, gfc_expr *expr, int parity)
 	 as 'long long'.  */
       gcc_assert (argsize == 2 * LONG_LONG_TYPE_SIZE);
 
-      func = builtin_decl_explicit (parity
-				    ? BUILT_IN_PARITYLL
-				    : BUILT_IN_POPCOUNTLL);
-
       /* Convert it to an integer, and store into a variable.  */
       utype = gfc_build_uint_type (argsize);
       arg = fold_convert (utype, arg);
       arg = gfc_evaluate_now (arg, &se->pre);
-
-      /* Call the builtin twice.  */
-      call1 = build_call_expr_loc (input_location, func, 1,
-				   fold_convert (long_long_unsigned_type_node,
-						 arg));
-
-      arg2 = fold_build2_loc (input_location, RSHIFT_EXPR, utype, arg,
-			      build_int_cst (utype, LONG_LONG_TYPE_SIZE));
-      call2 = build_call_expr_loc (input_location, func, 1,
-				   fold_convert (long_long_unsigned_type_node,
-						 arg2));
+      arg_type = long_long_unsigned_type_node;
 
       /* Combine the results.  */
       if (parity)
-	se->expr = fold_build2_loc (input_location, BIT_XOR_EXPR, result_type,
-				    call1, call2);
+	{
+	  /* Construct parityll (LOPART (arg) ^ HIPART (arg)) */
+	  arg2 = fold_build2_loc (input_location, RSHIFT_EXPR, utype, arg,
+				  build_int_cst (utype, LONG_LONG_TYPE_SIZE));
+	  arg = fold_build2_loc (input_location, BIT_XOR_EXPR, arg_type,
+				 fold_convert (arg_type, arg),
+				 fold_convert (arg_type, arg2));
+	  func = builtin_decl_explicit (BUILT_IN_PARITYLL);
+	  argsize = LONG_LONG_TYPE_SIZE;
+	}
       else
-	se->expr = fold_build2_loc (input_location, PLUS_EXPR, result_type,
-				    call1, call2);
-
-      return;
+	{
+	  func = builtin_decl_explicit (BUILT_IN_POPCOUNTLL);
+
+	  /* Call the builtin twice.  */
+	  call1 = build_call_expr_loc (input_location, func, 1,
+				       fold_convert (arg_type, arg));
+	  arg2 = fold_build2_loc (input_location, RSHIFT_EXPR, utype, arg,
+				  build_int_cst (utype, LONG_LONG_TYPE_SIZE));
+	  call2 = build_call_expr_loc (input_location, func, 1,
+				       fold_convert (arg_type, arg2));
+	  se->expr = fold_build2_loc (input_location, PLUS_EXPR, result_type,
+				      call1, call2);
+	  return;
+	}
     }
 
   /* Convert the actual argument twice: first, to the unsigned type of the

Re: [PATCH] gfortran: Improve translation of POPPAR intrinsic

[Bernhard Reutner-Fischer]

[-- Attachment #1: Type: text/plain, Size: 1592 bytes --]

Roger pinged this on gcc-patches some time ago fwiw.
[The commit-hooks will likely fix or ignore s/bext/next/ in his
mail-addr]


On Sun, 14 Jun 2020 23:39:32 +0100
"Roger Sayle" <roger@nextmovesoftware.com> wrote:

>  
> 
> The following patch to gfortran's trans-instrinsic.c tweaks the generic that
> is produced
> 
> for popcnt on integer(kind=16).  Currently, the double word popcnt is
> implemented as
> 
> parityll(hipart(x))^parityll(lopart(x)), but with this patch this is now
> translated as
> 
> parityll(hipart(x)^lopart(x)).  This will be just an aesthetic change once
> my tree-level
> 
> parity optimization patch of 12th June is reviewed and accepted, but
> generating the
> 
> more efficient form initially, avoids a tiny bit of garbage collection when
> the middle-end
> 
> cleans this up into its preferred form.    The semantics/correctness of this
> 
> change are tested by the run-time tests in gfortran.dg/popcnt_poppar_2.F90
> 
>  
> 
> This patch has been tested with "make bootstrap" and "make -k check" on
> 
> x86_64-pc-linux-gnu with no regressions.  If approved, I'd very much
> 
> appreciate it if the (gfortran) reviewer could commit this change for me.
> 
>  
> 
> 2020-06-14  Roger Sayle  <roger@bextmovesoftware.com>
> 
>  
> 
>         * trans-intrinsic.c (gfc_conv_intrinsic_popcnt_poppar): Translate
> 
>         poppar(kind=16) as parityll(hipart(x)^lopart(x)) instead of
> 
>         parityll(hipart(x))^parityll(lopart(x)).
> 
>  
> 
>  
> 
> Thanks in advance,
> 
> Roger
> 
> --
> 
> Roger Sayle
> 
> NextMove Software
> 
> Cambridge, UK
> 
>  
> 


[-- Attachment #2: patchb.txt --]
[-- Type: text/plain, Size: 2453 bytes --]

diff --git a/gcc/fortran/trans-intrinsic.c b/gcc/fortran/trans-intrinsic.c
index fd88099..363874e 100644
--- a/gcc/fortran/trans-intrinsic.c
+++ b/gcc/fortran/trans-intrinsic.c
@@ -7153,35 +7153,39 @@ gfc_conv_intrinsic_popcnt_poppar (gfc_se * se, gfc_expr *expr, int parity)
 	 as 'long long'.  */
       gcc_assert (argsize == 2 * LONG_LONG_TYPE_SIZE);
 
-      func = builtin_decl_explicit (parity
-				    ? BUILT_IN_PARITYLL
-				    : BUILT_IN_POPCOUNTLL);
-
       /* Convert it to an integer, and store into a variable.  */
       utype = gfc_build_uint_type (argsize);
       arg = fold_convert (utype, arg);
       arg = gfc_evaluate_now (arg, &se->pre);
-
-      /* Call the builtin twice.  */
-      call1 = build_call_expr_loc (input_location, func, 1,
-				   fold_convert (long_long_unsigned_type_node,
-						 arg));
-
-      arg2 = fold_build2_loc (input_location, RSHIFT_EXPR, utype, arg,
-			      build_int_cst (utype, LONG_LONG_TYPE_SIZE));
-      call2 = build_call_expr_loc (input_location, func, 1,
-				   fold_convert (long_long_unsigned_type_node,
-						 arg2));
+      arg_type = long_long_unsigned_type_node;
 
       /* Combine the results.  */
       if (parity)
-	se->expr = fold_build2_loc (input_location, BIT_XOR_EXPR, result_type,
-				    call1, call2);
+	{
+	  /* Construct parityll (LOPART (arg) ^ HIPART (arg)) */
+	  arg2 = fold_build2_loc (input_location, RSHIFT_EXPR, utype, arg,
+				  build_int_cst (utype, LONG_LONG_TYPE_SIZE));
+	  arg = fold_build2_loc (input_location, BIT_XOR_EXPR, arg_type,
+				 fold_convert (arg_type, arg),
+				 fold_convert (arg_type, arg2));
+	  func = builtin_decl_explicit (BUILT_IN_PARITYLL);
+	  argsize = LONG_LONG_TYPE_SIZE;
+	}
       else
-	se->expr = fold_build2_loc (input_location, PLUS_EXPR, result_type,
-				    call1, call2);
-
-      return;
+	{
+	  func = builtin_decl_explicit (BUILT_IN_POPCOUNTLL);
+
+	  /* Call the builtin twice.  */
+	  call1 = build_call_expr_loc (input_location, func, 1,
+				       fold_convert (arg_type, arg));
+	  arg2 = fold_build2_loc (input_location, RSHIFT_EXPR, utype, arg,
+				  build_int_cst (utype, LONG_LONG_TYPE_SIZE));
+	  call2 = build_call_expr_loc (input_location, func, 1,
+				       fold_convert (arg_type, arg2));
+	  se->expr = fold_build2_loc (input_location, PLUS_EXPR, result_type,
+				      call1, call2);
+	  return;
+	}
     }
 
   /* Convert the actual argument twice: first, to the unsigned type of the

Re: [PATCH] gfortran: Improve translation of POPPAR intrinsic

[Harald Anlauf]

Let's have a look at the tree-dump of the existing testcase:

integer(kind=4) runtime_poppar (integer(kind=16) & restrict i)
{
   integer(kind=4) res;

   {
     uint128_t D.4221;

     D.4221 = (uint128_t) *i;
     res = __builtin_parityll ((unsigned long) D.4221 ^ (unsigned long)
(D.4221 >> 64));
   }
   return res;
}

My understanding is there is actually nothing left to do,
as the middle-end(?) already handles this.

Am 21.11.21 um 01:22 schrieb Bernhard Reutner-Fischer via Fortran:
> Roger pinged this on gcc-patches some time ago fwiw.
> [The commit-hooks will likely fix or ignore s/bext/next/ in his
> mail-addr]
>
>
> On Sun, 14 Jun 2020 23:39:32 +0100
> "Roger Sayle" <roger@nextmovesoftware.com> wrote:
>
>>
>>
>> The following patch to gfortran's trans-instrinsic.c tweaks the generic that
>> is produced
>>
>> for popcnt on integer(kind=16).  Currently, the double word popcnt is
>> implemented as
>>
>> parityll(hipart(x))^parityll(lopart(x)), but with this patch this is now
>> translated as
>>
>> parityll(hipart(x)^lopart(x)).  This will be just an aesthetic change once
>> my tree-level
>>
>> parity optimization patch of 12th June is reviewed and accepted, but
>> generating the
>>
>> more efficient form initially, avoids a tiny bit of garbage collection when
>> the middle-end
>>
>> cleans this up into its preferred form.    The semantics/correctness of this
>>
>> change are tested by the run-time tests in gfortran.dg/popcnt_poppar_2.F90
>>
>>
>>
>> This patch has been tested with "make bootstrap" and "make -k check" on
>>
>> x86_64-pc-linux-gnu with no regressions.  If approved, I'd very much
>>
>> appreciate it if the (gfortran) reviewer could commit this change for me.
>>
>>
>>
>> 2020-06-14  Roger Sayle  <roger@bextmovesoftware.com>
>>
>>
>>
>>          * trans-intrinsic.c (gfc_conv_intrinsic_popcnt_poppar): Translate
>>
>>          poppar(kind=16) as parityll(hipart(x)^lopart(x)) instead of
>>
>>          parityll(hipart(x))^parityll(lopart(x)).
>>
>>
>>
>>
>>
>> Thanks in advance,
>>
>> Roger
>>
>> --
>>
>> Roger Sayle
>>
>> NextMove Software
>>
>> Cambridge, UK
>>
>>
>>
>

Re: [PATCH] gfortran: Improve translation of POPPAR intrinsic

[Bernhard Reutner-Fischer]

On Sun, 21 Nov 2021 19:59:35 +0100
Harald Anlauf <anlauf@gmx.de> wrote:

> Let's have a look at the tree-dump of the existing testcase:
> 
> integer(kind=4) runtime_poppar (integer(kind=16) & restrict i)
> {
>    integer(kind=4) res;
> 
>    {
>      uint128_t D.4221;
> 
>      D.4221 = (uint128_t) *i;
>      res = __builtin_parityll ((unsigned long) D.4221 ^ (unsigned long)
> (D.4221 >> 64));
>    }
>    return res;
> }
> 
> My understanding is there is actually nothing left to do,
> as the middle-end(?) already handles this.

Well the whole point was ...
> 
> Am 21.11.21 um 01:22 schrieb Bernhard Reutner-Fischer via Fortran:
> > Roger pinged this on gcc-patches some time ago fwiw.
> > [The commit-hooks will likely fix or ignore s/bext/next/ in his
> > mail-addr]
> >
> >
> > On Sun, 14 Jun 2020 23:39:32 +0100
> > "Roger Sayle" <roger@nextmovesoftware.com> wrote:
> >  
> >>
> >>
> >> The following patch to gfortran's trans-instrinsic.c tweaks the generic that
> >> is produced
> >>
> >> for popcnt on integer(kind=16).  Currently, the double word popcnt is
> >> implemented as
> >>
> >> parityll(hipart(x))^parityll(lopart(x)), but with this patch this is now
> >> translated as
> >>
> >> parityll(hipart(x)^lopart(x)).  This will be just an aesthetic change once
> >> my tree-level
> >>
> >> parity optimization patch of 12th June is reviewed and accepted, but
> >> generating the
> >>
> >> more efficient form initially, avoids a tiny bit of garbage collection when
> >> the middle-end
> >>
> >> cleans this up into its preferred form.    The semantics/correctness of this

... the above, i.e. create better code in the first place.
thanks,

> >>
> >> change are tested by the run-time tests in gfortran.dg/popcnt_poppar_2.F90
> >>
> >>
> >>
> >> This patch has been tested with "make bootstrap" and "make -k check" on
> >>
> >> x86_64-pc-linux-gnu with no regressions.  If approved, I'd very much
> >>
> >> appreciate it if the (gfortran) reviewer could commit this change for me.
> >>
> >>
> >>
> >> 2020-06-14  Roger Sayle  <roger@bextmovesoftware.com>
> >>
> >>
> >>
> >>          * trans-intrinsic.c (gfc_conv_intrinsic_popcnt_poppar): Translate
> >>
> >>          poppar(kind=16) as parityll(hipart(x)^lopart(x)) instead of
> >>
> >>          parityll(hipart(x))^parityll(lopart(x)).
> >>
> >>
> >>
> >>
> >>
> >> Thanks in advance,
> >>
> >> Roger
> >>
> >> --
> >>
> >> Roger Sayle
> >>
> >> NextMove Software
> >>
> >> Cambridge, UK
> >>
> >>
> >>  
> >  
>