[PATCH] Add constant folding for _mm*movemask* intrinsics (PR target/85317)

[Jakub Jelinek]

Hi!

The following patch handles constant folding of the builtins used in
*movemask* intrinsics - they have single operand and the only useful folding
seems to be if the argument is VECTOR_CST, we can do what the instruction
would do on that input and return the resulting INTEGER_CST.

Bootstrapped/regtested on x86_64-linux and i686-linux, ok for trunk?

2018-05-06  Jakub Jelinek  <jakub@redhat.com>

	PR target/85317
	* config/i386/i386.c (ix86_fold_builtin): Handle
	IX86_BUILTIN_{,P}MOVMSK{PS,PD,B}{,128,256}.

	* gcc.target/i386/pr85317.c: New test.
	* gcc.target/i386/avx2-vpmovmskb-2.c (avx2_test): Add asm volatile
	optimization barrier to avoid optimizing away the expected insn.

--- gcc/config/i386/i386.c.jj	2018-05-01 12:18:01.256814924 +0200
+++ gcc/config/i386/i386.c	2018-05-06 13:56:18.647354123 +0200
@@ -33477,6 +33477,37 @@ ix86_fold_builtin (tree fndecl, int n_ar
 	    }
 	  break;
 
+	case IX86_BUILTIN_MOVMSKPS:
+	case IX86_BUILTIN_PMOVMSKB:
+	case IX86_BUILTIN_MOVMSKPD:
+	case IX86_BUILTIN_PMOVMSKB128:
+	case IX86_BUILTIN_MOVMSKPD256:
+	case IX86_BUILTIN_MOVMSKPS256:
+	case IX86_BUILTIN_PMOVMSKB256:
+	  gcc_assert (n_args == 1);
+	  if (TREE_CODE (args[0]) == VECTOR_CST)
+	    {
+	      HOST_WIDE_INT res = 0;
+	      for (unsigned i = 0; i < VECTOR_CST_NELTS (args[0]); ++i)
+		{
+		  tree e = VECTOR_CST_ELT (args[0], i);
+		  if (TREE_CODE (e) == INTEGER_CST && !TREE_OVERFLOW (e))
+		    {
+		      if (wi::neg_p (wi::to_wide (e)))
+			res |= HOST_WIDE_INT_1 << i;
+		    }
+		  else if (TREE_CODE (e) == REAL_CST && !TREE_OVERFLOW (e))
+		    {
+		      if (TREE_REAL_CST (e).sign)
+			res |= HOST_WIDE_INT_1 << i;
+		    }
+		  else
+		    return NULL_TREE;
+		}
+	      return build_int_cst (TREE_TYPE (TREE_TYPE (fndecl)), res);
+	    }
+	  break;
+
 	default:
 	  break;
 	}
--- gcc/testsuite/gcc.target/i386/pr85317.c.jj	2018-05-06 14:28:59.201755826 +0200
+++ gcc/testsuite/gcc.target/i386/pr85317.c	2018-05-06 14:28:47.165747887 +0200
@@ -0,0 +1,38 @@
+/* PR target/85317 */
+/* { dg-do compile } */
+/* { dg-options "-O2 -mavx2 -fdump-tree-optimized" } */
+/* { dg-final { scan-tree-dump-not "link_error" "optimized" } } */
+
+#include <x86intrin.h>
+
+extern void link_error (void);
+
+int
+main ()
+{
+  int a = _mm_movemask_pd (_mm_set_pd (-2.0, 2.0));
+  if (a != 2) link_error ();
+  int b = _mm_movemask_pd (_mm_set_pd (0.0, __builtin_copysign (0.0, -4.0)));
+  if (b != 1) link_error ();
+  int c = _mm_movemask_ps (_mm_set_ps (__builtin_copysignf (0.0f, -4.0f), 0.0f,
+				       -4.0f, 4.0f));
+  if (c != 10) link_error ();
+  int d = _mm_movemask_epi8 (_mm_set_epi8 (-4, 8, -8, -12, 12, 15, 0, -1, -3,
+					   -128, 127, 126, 120, -120, 0, 5));
+  if (d != 0xb1c4) link_error ();
+  int e = _mm256_movemask_pd (_mm256_set_pd (-4.0, 0.0, 4.0,
+					     __builtin_copysign (0.0, -4.0)));
+  if (e != 9) link_error ();
+  int f = _mm256_movemask_ps (_mm256_set_ps (-8.0f, -16.0f, 12.0f, 127.0f,
+					     -4.0f, 0.0f, 4.0f,
+					     __builtin_copysign (0.0f,
+								 -4.0f)));
+  if (f != 0xc9) link_error ();
+  int g = _mm256_movemask_epi8 (_mm256_set_epi8 (-4, 8, -8, -12, 12, 15, 0, -1,
+						 -3, -128, 127, 126, 120, -120,
+						 0, 5, 12, 100, -20, -50, -70,
+						 2, -65, 0, -1, 1, 2, -2, -9,
+						 -9, 19, -64));
+  if (g != (int) 0xb1c43a9dU) link_error ();
+  return 0;
+}
--- gcc/testsuite/gcc.target/i386/avx2-vpmovmskb-2.c.jj	2016-05-22 12:20:18.514840455 +0200
+++ gcc/testsuite/gcc.target/i386/avx2-vpmovmskb-2.c	2018-05-07 09:16:22.766301274 +0200
@@ -14,6 +14,7 @@ avx2_test (void)
   s.x = _mm256_set_epi8 (1, 2, 3, 4, 10, 20, 30, 90, -80, -40, -100,
 			 15, 98, 25, 98, 7, 1, 2, 3, 4, 10, 20, 30, 90,
 			 -80, -40, -100, -15, 98, 25, 98, 7);
+  __asm volatile ("" : "+m" (s) : : "memory");
 
   res = _mm256_movemask_epi8 (s.x);
 

	Jakub