Li Jia He <heli...@linux.ibm.com> writes:
> Hi,
> @@ -546,6 +546,12 @@ DEFINE_INT_AND_FLOAT_ROUND_FN (RINT)
> (simplify
> (mod (mod@2 @0 @1) @1)
> @2)
> + /* Optimize (x shift (n mod C)) to (x shift (n bit_and (C - 1))) when C is a
> + constant and power of two. */
> + (for shift (lshift rshift)
> + (simplify
> + (shift @0 (mod @1 integer_pow2p@2))
> + (shift @0 (bit_and @1 (minus @2 { build_int_cst (TREE_TYPE (@2), 1);
> })))))
> /* From extract_muldiv_1: (X * C1) % C2 is zero if C1 is a multiple of C2.
> */
> (simplify
> (mod (mult @0 INTEGER_CST@1) INTEGER_CST@2)
> +unsigned a(unsigned x, int n)
> +{
> + return x >> (n % 32);
> +}
> +
With this patch, the behavior will change on negtive @2. While this
would not be an issue because it is UB.
In below foo, r and r1 are not same after patch.
---------
int __attribute__ ((noinline))
foo (unsigned x, int n, int t)
{
int r = x >> (n % 4);
int r1 = x >> t;
__builtin_printf ("%d : %d\n", r, r1);
return r == r1;
}
int main()
{
unsigned x = 0x12c;
int n = -3;
return foo (x, n, n %4);
}
---------
For: "x=0x12c, n=-3; x >> (n % 4)" without patch, it is 0; with patch
it is 0x96.
