My recent wide_int_binop changes caused ICE on this testcase.
The problem is that for shift where amount has MSB set now fails to optimize
into a constant (IMHO we should treat out of bounds shifts the same later),
but there is a precedent for that already - e.g. division by zero fails
to optimize into a constant too. I think it is better if path isolation
checks for these UBs and does something the user chooses (__builtin_trap vs.
__builtin_unreachable, and either a deferred warning about the UB or
nothing).
This patch just doesn't optimize if int_const_binop failed.
2020-11-26 Jakub Jelinek <jakub@redhat.com>
PR tree-optimization/97979
* match.pd ((X {&,^,|} C2) << C1 into (X << C1) {&,^,|} (C2 << C1)):
Only optimize if int_const_binop returned non-NULL.
* gcc.dg/pr97979.c: New test.
* gcc.c-torture/compile/pr97979.c: New test.
(shift (convert?:s (bit_op:s @0 INTEGER_CST@2)) INTEGER_CST@1)
(if (tree_nop_conversion_p (type, TREE_TYPE (@0)))
(with { tree mask = int_const_binop (shift, fold_convert (type, @2), @1); }
- (bit_op (shift (convert @0) @1) { mask; }))))))
+ (if (mask)
+ (bit_op (shift (convert @0) @1) { mask; })))))))
/* ~(~X >> Y) -> X >> Y (for arithmetic shift). */
(simplify
--- /dev/null
+/* PR tree-optimization/97979 */
+
+int
+foo (int x)
+{
+ return (x & 0x123) << -3;
+}
--- /dev/null
+/* PR tree-optimization/97979 */
+/* { dg-do compile } */
+/* { dg-options "-O2 -fno-tree-ccp" } */
+
+short a = 0;
+int b = 0;
+
+void
+foo (void)
+{
+ unsigned short d = b;
+ a = d >> -2U; /* { dg-warning "right shift count >= width of type" } */
+}