The following fixes wrong-code when SLP vectorizing shifts where we
may end up detecting the shift amount as scalar even though it
really isn't.
Bootstrapped and tested on x86_64-unknown-linux-gnu, applied on trunk
sofar.
Richard.
2019-01-18 Richard Biener <rguent...@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.
Index: gcc/testsuite/gcc.dg/vect/pr88903-1.c
===================================================================
--- gcc/testsuite/gcc.dg/vect/pr88903-1.c (nonexistent)
+++ gcc/testsuite/gcc.dg/vect/pr88903-1.c (working copy)
@@ -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;
+}
Index: gcc/testsuite/gcc.dg/vect/pr88903-2.c
===================================================================
--- gcc/testsuite/gcc.dg/vect/pr88903-2.c (nonexistent)
+++ gcc/testsuite/gcc.dg/vect/pr88903-2.c (working copy)
@@ -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;
+}
Index: gcc/tree-vect-stmts.c
===================================================================
--- gcc/tree-vect-stmts.c (revision 268068)
+++ gcc/tree-vect-stmts.c (working copy)
@@ -5540,6 +5540,15 @@ vectorizable_shift (stmt_vec_info stmt_i
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
