Hello,

this is another small simplification of RTL for vectors. Note that it doesn't really solve the problem, because these simplifications are only performed for single-use objects. If I start from vectors [a,b] and [c,d] and concatenate them into [a,b,c,d], then extract both halves, as in the original testcase in the PR, we won't notice that those are the original vectors. Still, better than nothing...

(we output a vzeroupper for the testcase, that seems unnecessary)

Bootstrap+testsuite on x86_64-linux-gnu.

2014-06-22  Marc Glisse  <marc.gli...@inria.fr>

        PR target/44551
gcc/
        * simplify-rtx.c (simplify_binary_operation_1) <VEC_SELECT>:
        Optimize inverse of a VEC_CONCAT.
gcc/testsuite/
        * gcc.target/i386/pr44551-1.c: New file.

--
Marc Glisse
Index: gcc/simplify-rtx.c
===================================================================
--- gcc/simplify-rtx.c  (revision 211867)
+++ gcc/simplify-rtx.c  (working copy)
@@ -3359,20 +3359,64 @@ simplify_binary_operation_1 (enum rtx_co
              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 we select one half of a vec_concat, return that.  */
+         if (GET_CODE (trueop0) == VEC_CONCAT
+             && CONST_INT_P (XVECEXP (trueop1, 0, 0)))
+           {
+             rtx subop0 = XEXP (trueop0, 0);
+             rtx subop1 = XEXP (trueop0, 1);
+             enum machine_mode mode0 = GET_MODE (subop0);
+             enum machine_mode mode1 = GET_MODE (subop1);
+             int li = GET_MODE_SIZE (GET_MODE_INNER (mode0));
+             int l0 = GET_MODE_SIZE (mode0) / li;
+             int l1 = GET_MODE_SIZE (mode1) / li;
+             int i0 = INTVAL (XVECEXP (trueop1, 0, 0));
+             if (i0 == 0 && !side_effects_p (op1) && mode == mode0)
+               {
+                 bool success = true;
+                 for (int i = 1; i < l0; ++i)
+                   {
+                     rtx j = XVECEXP (trueop1, 0, i);
+                     if (!CONST_INT_P (j) || INTVAL (j) != i)
+                       {
+                         success = false;
+                         break;
+                       }
+                   }
+                 if (success)
+                   return subop0;
+               }
+             if (i0 == l0 && !side_effects_p (op0) && mode == mode1)
+               {
+                 bool success = true;
+                 for (int i = 1; i < l1; ++i)
+                   {
+                     rtx j = XVECEXP (trueop1, 0, i);
+                     if (!CONST_INT_P (j) || INTVAL (j) != i0 + i)
+                       {
+                         success = false;
+                         break;
+                       }
+                   }
+                 if (success)
+                   return 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.  */
Index: gcc/testsuite/gcc.target/i386/pr44551-1.c
===================================================================
--- gcc/testsuite/gcc.target/i386/pr44551-1.c   (revision 0)
+++ gcc/testsuite/gcc.target/i386/pr44551-1.c   (working copy)
@@ -0,0 +1,15 @@
+/* { dg-do compile } */
+/* { dg-options "-O2 -mavx" } */
+
+#include <immintrin.h>
+
+__m128i
+foo (__m256i x, __m128i y)
+{
+  __m256i r = _mm256_insertf128_si256(x, y, 1);
+  __m128i a = _mm256_extractf128_si256(r, 1);
+  return a;
+}
+
+/* { dg-final { scan-assembler-not "vinsertf" } } */
+/* { dg-final { scan-assembler-not "vextractf" } } */

Reply via email to