Re: [Qemu-devel] [PATCH v2 3/4] tcg: Mask shift counts to avoid undefined behavior

[Stefan Weil]

Am 18.03.2014 22:30, schrieb Richard Henderson:
> TCG now requires unspecified behavior rather than a potential crash,
> bring the C shift within the letter of the law.

I know that C does not define the result of some shift / rotate
operations, but I don't understand the sentence above. Why does TCG or
TCI require unspecified behaviour now? Where was or is a potential crash?

The modifications below won't harm, but make the TCG interpreter slower.
Are they (all) necessary? Are there test cases which fail with the old code?

> 
> Signed-off-by: Richard Henderson <address@hidden>
> ---
>  tci.c | 20 ++++++++++----------
>  1 file changed, 10 insertions(+), 10 deletions(-)
> 
> diff --git a/tci.c b/tci.c
> index 0202ed9..6523ab8 100644
> --- a/tci.c
> +++ b/tci.c
> @@ -669,32 +669,32 @@ uintptr_t tcg_qemu_tb_exec(CPUArchState *env, uint8_t 
> *tb_ptr)
>              t0 = *tb_ptr++;
>              t1 = tci_read_ri32(&tb_ptr);
>              t2 = tci_read_ri32(&tb_ptr);
> -            tci_write_reg32(t0, t1 << t2);
> +            tci_write_reg32(t0, t1 << (t2 & 31));
>              break;
>          case INDEX_op_shr_i32:
>              t0 = *tb_ptr++;
>              t1 = tci_read_ri32(&tb_ptr);
>              t2 = tci_read_ri32(&tb_ptr);
> -            tci_write_reg32(t0, t1 >> t2);
> +            tci_write_reg32(t0, t1 >> (t2 & 31));

Right shifts of unsigned values with unsigned shift count are always
defined, aren't they? So masking for those cases should not be needed.

>              break;
>          case INDEX_op_sar_i32:
>              t0 = *tb_ptr++;
>              t1 = tci_read_ri32(&tb_ptr);
>              t2 = tci_read_ri32(&tb_ptr);
> -            tci_write_reg32(t0, ((int32_t)t1 >> t2));
> +            tci_write_reg32(t0, ((int32_t)t1 >> (t2 & 31)));
>              break;
>  #if TCG_TARGET_HAS_rot_i32
>          case INDEX_op_rotl_i32:
>              t0 = *tb_ptr++;
>              t1 = tci_read_ri32(&tb_ptr);
>              t2 = tci_read_ri32(&tb_ptr);
> -            tci_write_reg32(t0, rol32(t1, t2));
> +            tci_write_reg32(t0, rol32(t1, t2 & 31));

What about other users of rol32?

>              break;
>          case INDEX_op_rotr_i32:
>              t0 = *tb_ptr++;
>              t1 = tci_read_ri32(&tb_ptr);
>              t2 = tci_read_ri32(&tb_ptr);
> -            tci_write_reg32(t0, ror32(t1, t2));
> +            tci_write_reg32(t0, ror32(t1, t2 & 31));
>              break;
>  #endif
>  #if TCG_TARGET_HAS_deposit_i32
> @@ -936,32 +936,32 @@ uintptr_t tcg_qemu_tb_exec(CPUArchState *env, uint8_t 
> *tb_ptr)
>              t0 = *tb_ptr++;
>              t1 = tci_read_ri64(&tb_ptr);
>              t2 = tci_read_ri64(&tb_ptr);
> -            tci_write_reg64(t0, t1 << t2);
> +            tci_write_reg64(t0, t1 << (t2 & 63));
>              break;
>          case INDEX_op_shr_i64:
>              t0 = *tb_ptr++;
>              t1 = tci_read_ri64(&tb_ptr);
>              t2 = tci_read_ri64(&tb_ptr);
> -            tci_write_reg64(t0, t1 >> t2);
> +            tci_write_reg64(t0, t1 >> (t2 & 63));
>              break;
>          case INDEX_op_sar_i64:
>              t0 = *tb_ptr++;
>              t1 = tci_read_ri64(&tb_ptr);
>              t2 = tci_read_ri64(&tb_ptr);
> -            tci_write_reg64(t0, ((int64_t)t1 >> t2));
> +            tci_write_reg64(t0, ((int64_t)t1 >> (t2 & 63)));
>              break;
>  #if TCG_TARGET_HAS_rot_i64
>          case INDEX_op_rotl_i64:
>              t0 = *tb_ptr++;
>              t1 = tci_read_ri64(&tb_ptr);
>              t2 = tci_read_ri64(&tb_ptr);
> -            tci_write_reg64(t0, rol64(t1, t2));
> +            tci_write_reg64(t0, rol64(t1, t2 & 63));
>              break;
>          case INDEX_op_rotr_i64:
>              t0 = *tb_ptr++;
>              t1 = tci_read_ri64(&tb_ptr);
>              t2 = tci_read_ri64(&tb_ptr);
> -            tci_write_reg64(t0, ror64(t1, t2));
> +            tci_write_reg64(t0, ror64(t1, t2 & 63));
>              break;
>  #endif
>  #if TCG_TARGET_HAS_deposit_i64
> 

Regards
Stefan