Re: [Qemu-devel] [PATCH v12 09/14] target-mips: Add ASE DSP bit/manipulation instructions

[Richard Henderson <rth@twiddle.net>]

On 2012-10-31 01:44, Peter Maydell wrote:
> On 30 October 2012 15:34, Jia Liu <proljc@gmail.com> wrote:
>> On Mon, Oct 29, 2012 at 9:40 PM, Jovanovic, Petar <petarj@mips.com> wrote:
>>>> imm = (int16_t)(imm << 6) >> 6;
>>>
>>> result of a bitwise shift of a signed type and a negative vlaue is
>>> implementation-defined, so you can not rely on that.
>>>
>>
>> I think it will take a 10bits signed value sign extend into 16bits
>> signed value, and I've tested it with negative values, it working
>> well.
> 
> You cannot rely on the behaviour of a specific compiler implementation
> as evidence that a piece of code is correct. C has a standard which
> defines what is and is not valid.

Indeed.  The only portable way is

  val = ((val & (sign | (sign - 1))) ^ sign) - sign

with all unsigned types, and "sign" set to the sign bit.

> 
> Having said that, right shift of negative signed integers is one of
> those bits of implementation defined behaviour which we allow ourselves
> to rely on in QEMU because all the platforms we care about behave
> that way. (That integers are 2s complement representation is another.)

Also very true.  I don't like seeing the code in question though.
We've several implementations of sign-extend-to-N-bits functions
throughout qemu; we ought to unify them.


r~