Re: [PATCH] tcg: Sanitize shift constants on ppc64le so that shift operations with large constants don't generate invalid instructions.

[Richard Henderson]

On 6/2/20 11:43 PM, Philippe Mathieu-Daudé wrote:
> 
> Hi Catherine,
> 
> On 6/3/20 7:23 AM, agrecascino123@gmail.com wrote:
>> From: "Catherine A. Frederick" <chocola@animebitch.es>
>>
>> Signed-off-by: "Catherine A. Frederick" <chocola@animebitch.es>
>> ---
>>  tcg/ppc/tcg-target.inc.c | 4 ++++
>>  1 file changed, 4 insertions(+)
>>
>> diff --git a/tcg/ppc/tcg-target.inc.c b/tcg/ppc/tcg-target.inc.c
>> index ee1f9227c1..a5450a5e67 100644
>> --- a/tcg/ppc/tcg-target.inc.c
>> +++ b/tcg/ppc/tcg-target.inc.c
>> @@ -790,21 +790,25 @@ static inline void tcg_out_ext32u(TCGContext *s, 
>> TCGReg dst, TCGReg src)
>>  
>>  static inline void tcg_out_shli32(TCGContext *s, TCGReg dst, TCGReg src, 
>> int c)
>>  {
>> +    c = ((unsigned)c > 32) ? 32 : c;
>>      tcg_out_rlw(s, RLWINM, dst, src, c, 0, 31 - c);
>>  }
>>  
>>  static inline void tcg_out_shli64(TCGContext *s, TCGReg dst, TCGReg src, 
>> int c)
>>  {
>> +    c = ((unsigned)c > 64) ? 64 : c;
>>      tcg_out_rld(s, RLDICR, dst, src, c, 63 - c);
>>  }
>>  
>>  static inline void tcg_out_shri32(TCGContext *s, TCGReg dst, TCGReg src, 
>> int c)
>>  {
>> +    c = ((unsigned)c > 32) ? 32 : c;
>>      tcg_out_rlw(s, RLWINM, dst, src, 32 - c, c, 31);
>>  }
>>  
>>  static inline void tcg_out_shri64(TCGContext *s, TCGReg dst, TCGReg src, 
>> int c)
>>  {
>> +    c = ((unsigned)c > 64) ? 64 : c;
>>      tcg_out_rld(s, RLDICL, dst, src, 64 - c, c);
>>  }
> 
> I agree there is a bug, but I am not sure we should silently cap the
> value this way. I'd rather see the caller provide a value in range, and
> maybe the callee use 'tcg_debug_assert(c <= RANGE);' to catch future new
> caller added missing the range check.

We have done this before: see 1fd95946657.

In tcg/README, we note that out-of-range shifts produce undefined results, but
should not trap with illegal instruction.

I would like to know more about where these out-of-range shifts are being
generated, but I do know that there are innocent ways by which this can happen.

For instance, one way in which we can translate a guest in which out-of-range
shifts produce zero is

  x = (shift < 32 ? y << shift : 0)

using INDEX_op_movcond_i32 for the ?: operator.  Which means that
we use the original (out-of-range) shift and subsequently discard the undefined
result.

Catherine, I think it would be more appropriate to mask C rather than bound it
to another out-of-range value: c &= 31 or c &= 64, with a comment about
avoiding an illegal instruction, just as in the tcg/sparc patch I reference 
above.


r~