On Fri, Sep 18, 2020 at 01:39:16PM +0200, Richard Biener wrote:
> --- a/gcc/tree-vect-patterns.c
> +++ b/gcc/tree-vect-patterns.c
> @@ -2456,7 +2456,6 @@ vect_recog_rotate_pattern (vec_info *vinfo,
> append_pattern_def_seq (vinfo, stmt_vinfo, def_stmt);
> }
> stype = TREE_TYPE (def);
> - scalar_int_mode smode = SCALAR_INT_TYPE_MODE (stype);
>
> if (TREE_CODE (def) == INTEGER_CST)
> {
> @@ -2485,7 +2484,7 @@ vect_recog_rotate_pattern (vec_info *vinfo,
> append_pattern_def_seq (vinfo, stmt_vinfo, def_stmt, vecstype);
>
> def2 = vect_recog_temp_ssa_var (stype, NULL);
> - tree mask = build_int_cst (stype, GET_MODE_PRECISION (smode) - 1);
> + tree mask = build_int_cst (stype, GET_MODE_PRECISION (mode) - 1);
> def_stmt = gimple_build_assign (def2, BIT_AND_EXPR,
> gimple_assign_lhs (def_stmt), mask);
> if (ext_def)
The bug didn't affect just the > 32 precision type rotates, but also the
narrower ones where we'd end up with UB on the vector shifts and some arches
like x86 would end up with 0 as the result of the shift rather than the
shift count being masked with the operand width.
So I think it is worth adding testcase for that too. Tested on x86_64-linux
-m32/-m64 without and with your fix, committed to trunk as obvious:
2020-09-18 Jakub Jelinek <ja...@redhat.com>
PR tree-optimization/97081
* gcc.dg/vect/pr97081-2.c: New test.
--- gcc/testsuite/gcc.dg/vect/pr97081-2.c
+++ gcc/testsuite/gcc.dg/vect/pr97081-2.c
@@ -0,0 +1,32 @@
+/* PR tree-optimization/97081 */
+
+#include "tree-vect.h"
+
+unsigned short s[1024];
+unsigned char c[1024];
+
+__attribute__((noipa)) void
+foo (int n)
+{
+ for (int i = 0; i < 1024; i++)
+ s[i] = (s[i] << n) | (s[i] >> (__SIZEOF_SHORT__ * __CHAR_BIT__ - n));
+ for (int i = 0; i < 1024; i++)
+ c[i] = (c[i] << n) | (c[i] >> (__CHAR_BIT__ - n));
+}
+
+int
+main ()
+{
+ check_vect ();
+ for (int i = 0; i < 1024; i++)
+ {
+ s[i] = i;
+ c[i] = i;
+ }
+ foo (3);
+ for (int i = 0; i < 1024; i++)
+ if (s[i] != (unsigned short) ((i << 3) | (i >> (__SIZEOF_SHORT__ *
__CHAR_BIT__ - 3)))
+ || c[i] != (unsigned char) ((((unsigned char) i) << 3) | (((unsigned
char) i) >> (__CHAR_BIT__ - 3))))
+ __builtin_abort ();
+ return 0;
+}
Jakub
