re PR tree-optimization/88903 (wrong-code with SLP vectorized shift)

2019-01-18  Richard Biener  <rguenther@suse.de>

	PR tree-optimization/88903
	* tree-vect-stmts.c (vectorizable_shift): Verify we see all
	scalar stmts a SLP shift amount is composed of when detecting
	shifts by scalars.

	* gcc.dg/vect/pr88903-1.c: New testcase.
	* gcc.dg/vect/pr88903-2.c: Likewise.

From-SVN: r268076

5 files changed, 77 insertions, 0 deletions

diff --git a/gcc/ChangeLog b/gcc/ChangeLog
index ff87cbe..1389077 100644
--- a/gcc/ChangeLog
+++ b/gcc/ChangeLog
@@ -1,3 +1,10 @@
+2019-01-18  Richard Biener  <rguenther@suse.de>
+
+	PR tree-optimization/88903
+	* tree-vect-stmts.c (vectorizable_shift): Verify we see all
+	scalar stmts a SLP shift amount is composed of when detecting
+	shifts by scalars.
+
 2019-01-18  Richard Earnshaw  <rearnsha@arm.com>
 
 	PR target/88799
diff --git a/gcc/testsuite/ChangeLog b/gcc/testsuite/ChangeLog
index 88a590e..77b89f4 100644
--- a/gcc/testsuite/ChangeLog
+++ b/gcc/testsuite/ChangeLog
@@ -1,3 +1,9 @@
+2019-01-18  Richard Biener  <rguenther@suse.de>
+
+	PR tree-optimization/88903
+	* gcc.dg/vect/pr88903-1.c: New testcase.
+	* gcc.dg/vect/pr88903-2.c: Likewise.
+
 2019-01-18  H.J. Lu  <hongjiu.lu@intel.com>
 
 	PR c/51628
diff --git a/gcc/testsuite/gcc.dg/vect/pr88903-1.c b/gcc/testsuite/gcc.dg/vect/pr88903-1.c
new file mode 100644
index 0000000..dead2b5
--- /dev/null
+++ b/gcc/testsuite/gcc.dg/vect/pr88903-1.c
@@ -0,0 +1,26 @@
+#include "tree-vect.h"
+
+int x[1024];
+
+void __attribute__((noinline))
+foo()
+{
+  for (int i = 0; i < 512; ++i)
+    {
+      x[2*i] = x[2*i] << (i+1);
+      x[2*i+1] = x[2*i+1] << (i+1);
+    }
+}
+
+int
+main()
+{
+  check_vect ();
+  for (int i = 0; i < 1024; ++i)
+    x[i] = i;
+  foo ();
+  for (int i = 0; i < 1024; ++i)
+    if (x[i] != i << (i/2+1))
+      __builtin_abort ();
+  return 0;
+}
diff --git a/gcc/testsuite/gcc.dg/vect/pr88903-2.c b/gcc/testsuite/gcc.dg/vect/pr88903-2.c
new file mode 100644
index 0000000..cd88a99
--- /dev/null
+++ b/gcc/testsuite/gcc.dg/vect/pr88903-2.c
@@ -0,0 +1,28 @@
+#include "tree-vect.h"
+
+int x[1024];
+int y[1024];
+int z[1024];
+
+void __attribute__((noinline)) foo()
+{
+  for (int i = 0; i < 512; ++i)
+    {
+      x[2*i] = x[2*i] << y[2*i];
+      x[2*i+1] = x[2*i+1] << y[2*i];
+      z[2*i] = y[2*i];
+      z[2*i+1] = y[2*i+1];
+    }
+}
+
+int main()
+{
+  check_vect ();
+  for (int i = 0; i < 1024; ++i)
+    x[i] = i, y[i] = i % 8;
+  foo ();
+  for (int i = 0; i < 1024; ++i)
+    if (x[i] != i << ((i & ~1) % 8))
+      __builtin_abort ();
+  return 0;
+}
diff --git a/gcc/tree-vect-stmts.c b/gcc/tree-vect-stmts.c
index c0e19dd..2220fce 100644
--- a/gcc/tree-vect-stmts.c
+++ b/gcc/tree-vect-stmts.c
@@ -5540,6 +5540,16 @@ vectorizable_shift (stmt_vec_info stmt_info, gimple_stmt_iterator *gsi,
 	      if (!operand_equal_p (gimple_assign_rhs2 (slpstmt), op1, 0))
 		scalar_shift_arg = false;
 	    }
+
+	  /* For internal SLP defs we have to make sure we see scalar stmts
+	     for all vector elements.
+	     ???  For different vectors we could resort to a different
+	     scalar shift operand but code-generation below simply always
+	     takes the first.  */
+	  if (dt[1] == vect_internal_def
+	      && maybe_ne (nunits_out * SLP_TREE_NUMBER_OF_VEC_STMTS (slp_node),
+			   stmts.length ()))
+	    scalar_shift_arg = false;
 	}
 
       /* If the shift amount is computed by a pattern stmt we cannot