[rtl, i386] combine concat+permutation

[rtl, i386] combine concat+permutation

[Marc Glisse]

[-- Attachment #1: Type: TEXT/PLAIN, Size: 474 bytes --]

Hello,

this is a follow up to the patch applied after this discussion:
http://gcc.gnu.org/ml/gcc-patches/2012-05/msg00504.html

It handles the -mavx __builtin_shuffle case mentioned there.

It passes bootstrap (languages=c,c++) and regtest on x86_64.


2012-08-04  Marc Glisse  <marc.glisse@inria.fr>

gcc/
 	* simplify-rtx.c (simplify_binary_operation_1): Optimize shuffle of
 	a concatenation.

gcc/testsuite/
 	* gcc.target/i386/perm-concat.c: New test.

-- 
Marc Glisse

[-- Attachment #2: Type: TEXT/PLAIN, Size: 3074 bytes --]

Index: gcc/testsuite/gcc.target/i386/perm-concat.c
===================================================================
--- gcc/testsuite/gcc.target/i386/perm-concat.c	(revision 0)
+++ gcc/testsuite/gcc.target/i386/perm-concat.c	(revision 0)
@@ -0,0 +1,13 @@
+/* { dg-do compile } */
+/* { dg-options "-O -mavx -mfpmath=sse" } */
+
+typedef double v2df __attribute__ ((__vector_size__ (16)));
+
+v2df
+f (double d)
+{
+  v2df x = {-d, d};
+  return __builtin_ia32_vpermilpd (x, 1);
+}
+
+/* { dg-final { scan-assembler-not "\tvpermilpd\[ \t\]" } } */

Property changes on: gcc/testsuite/gcc.target/i386/perm-concat.c
___________________________________________________________________
Added: svn:keywords
   + Author Date Id Revision URL
Added: svn:eol-style
   + native

Index: gcc/simplify-rtx.c
===================================================================
--- gcc/simplify-rtx.c	(revision 190127)
+++ gcc/simplify-rtx.c	(working copy)
@@ -3235,40 +3235,57 @@ simplify_binary_operation_1 (enum rtx_co
 	  /* If we build {a,b} then permute it, build the result directly.  */
 	  if (XVECLEN (trueop1, 0) == 2
 	      && CONST_INT_P (XVECEXP (trueop1, 0, 0))
 	      && CONST_INT_P (XVECEXP (trueop1, 0, 1))
 	      && GET_CODE (trueop0) == VEC_CONCAT
 	      && GET_CODE (XEXP (trueop0, 0)) == VEC_CONCAT
 	      && GET_MODE (XEXP (trueop0, 0)) == mode
 	      && GET_CODE (XEXP (trueop0, 1)) == VEC_CONCAT
 	      && GET_MODE (XEXP (trueop0, 1)) == mode)
 	    {
 	      unsigned int i0 = INTVAL (XVECEXP (trueop1, 0, 0));
 	      unsigned int i1 = INTVAL (XVECEXP (trueop1, 0, 1));
 	      rtx subop0, subop1;
 
 	      gcc_assert (i0 < 4 && i1 < 4);
 	      subop0 = XEXP (XEXP (trueop0, i0 / 2), i0 % 2);
 	      subop1 = XEXP (XEXP (trueop0, i1 / 2), i1 % 2);
 
 	      return simplify_gen_binary (VEC_CONCAT, mode, subop0, subop1);
 	    }
+
+	  if (XVECLEN (trueop1, 0) == 2
+	      && CONST_INT_P (XVECEXP (trueop1, 0, 0))
+	      && CONST_INT_P (XVECEXP (trueop1, 0, 1))
+	      && GET_CODE (trueop0) == VEC_CONCAT
+	      && GET_MODE (trueop0) == mode)
+	    {
+	      unsigned int i0 = INTVAL (XVECEXP (trueop1, 0, 0));
+	      unsigned int i1 = INTVAL (XVECEXP (trueop1, 0, 1));
+	      rtx subop0, subop1;
+
+	      gcc_assert (i0 < 2 && i1 < 2);
+	      subop0 = XEXP (trueop0, i0);
+	      subop1 = XEXP (trueop0, i1);
+
+	      return simplify_gen_binary (VEC_CONCAT, mode, subop0, subop1);
+	    }
 	}
 
       if (XVECLEN (trueop1, 0) == 1
 	  && CONST_INT_P (XVECEXP (trueop1, 0, 0))
 	  && GET_CODE (trueop0) == VEC_CONCAT)
 	{
 	  rtx vec = trueop0;
 	  int offset = INTVAL (XVECEXP (trueop1, 0, 0)) * GET_MODE_SIZE (mode);
 
 	  /* Try to find the element in the VEC_CONCAT.  */
 	  while (GET_MODE (vec) != mode
 		 && GET_CODE (vec) == VEC_CONCAT)
 	    {
 	      HOST_WIDE_INT vec_size = GET_MODE_SIZE (GET_MODE (XEXP (vec, 0)));
 	      if (offset < vec_size)
 		vec = XEXP (vec, 0);
 	      else
 		{
 		  offset -= vec_size;
 		  vec = XEXP (vec, 1);

Re: [rtl, i386] combine concat+permutation

[Marc Glisse]

Ping
http://gcc.gnu.org/ml/gcc-patches/2012-08/msg00205.html

(Cc: the nice reviewer of the previous patch)

On Fri, 3 Aug 2012, Marc Glisse wrote:

> Hello,
>
> this is a follow up to the patch applied after this discussion:
> http://gcc.gnu.org/ml/gcc-patches/2012-05/msg00504.html
>
> It handles the -mavx __builtin_shuffle case mentioned there.
>
> It passes bootstrap (languages=c,c++) and regtest on x86_64.
>
>
> 2012-08-04  Marc Glisse  <marc.glisse@inria.fr>
>
> gcc/
> 	* simplify-rtx.c (simplify_binary_operation_1): Optimize shuffle of
> 	a concatenation.
>
> gcc/testsuite/
> 	* gcc.target/i386/perm-concat.c: New test.

-- 
Marc Glisse

Re: [rtl, i386] combine concat+permutation

[Jakub Jelinek]

On Fri, Aug 03, 2012 at 10:47:25PM +0200, Marc Glisse wrote:
> Hello,
> 
> this is a follow up to the patch applied after this discussion:
> http://gcc.gnu.org/ml/gcc-patches/2012-05/msg00504.html
> 
> It handles the -mavx __builtin_shuffle case mentioned there.
> 
> It passes bootstrap (languages=c,c++) and regtest on x86_64.

Ok.  Thanks.

> 2012-08-04  Marc Glisse  <marc.glisse@inria.fr>
> 
> gcc/
> 	* simplify-rtx.c (simplify_binary_operation_1): Optimize shuffle of
> 	a concatenation.
> 
> gcc/testsuite/
> 	* gcc.target/i386/perm-concat.c: New test.

	Jakub

Re: [rtl, i386] combine concat+permutation

[Marc Glisse]

On Fri, 17 Aug 2012, Jakub Jelinek wrote:

> On Fri, Aug 03, 2012 at 10:47:25PM +0200, Marc Glisse wrote:
>> Hello,
>>
>> this is a follow up to the patch applied after this discussion:
>> http://gcc.gnu.org/ml/gcc-patches/2012-05/msg00504.html
>>
>> It handles the -mavx __builtin_shuffle case mentioned there.
>>
>> It passes bootstrap (languages=c,c++) and regtest on x86_64.
>
> Ok.  Thanks.
>
>> 2012-08-04  Marc Glisse  <marc.glisse@inria.fr>
>>
>> gcc/
>> 	* simplify-rtx.c (simplify_binary_operation_1): Optimize shuffle of
>> 	a concatenation.
>>
>> gcc/testsuite/
>> 	* gcc.target/i386/perm-concat.c: New test.

Thanks, committed.

I just noticed that for __builtin_shuffle, this is an optimization that 
fold_ternary knows, but which never happens for 2 reasons:

1) the first time it sees the VEC_PERM_EXPR, the last argument is a 
CONSTRUCTOR but it only handles VECTOR_CST (this also breaks constexpr 
in C++11)

2) arg0 and arg1 are ssa names, not directly constructors, and there is no 
combine pass.

fold_ternary does get called quite a few times on it, so I tried cheating 
in gdb and replacing the ssa names by the constructors, and it did fold as 
expected, but the optimization was then discarded because 
valid_gimple_rhs_p doesn't like constructors.

I guess I'll have to add it to forwprop (it can make sense to restrict it 
to the case where the constructor is only used once, since shuffles might 
be faster than constructions).

-- 
Marc Glisse