[Mesa-dev] [PATCH 1/2] gm107/ir: make use of FADD32I for all immediates

[Samuel Pitoiset]

On 06/28/2016 12:06 AM, Ilia Mirkin wrote:
> On Mon, Jun 27, 2016 at 6:05 PM, Ilia Mirkin <imirkin at alum.mit.edu> wrote:
>> On Mon, Jun 27, 2016 at 6:04 PM, Samuel Pitoiset
>> <samuel.pitoiset at gmail.com> wrote:
>>>
>>>
>>> On 06/28/2016 12:02 AM, Ilia Mirkin wrote:
>>>>
>>>> This loses you saturation. Does the target account for this?
>>>
>>>
>>> No saturate flag for FADD32I.
>>
>> That's not what I asked.
>
> Specifically look at this code:
>
> bool
> TargetNVC0::isSatSupported(const Instruction *insn) const
> {
>    if (insn->op == OP_CVT)
>       return true;
>    if (!(opInfo[insn->op].dstMods & NV50_IR_MOD_SAT))
>       return false;
>
>    if (insn->dType == TYPE_U32)
>       return (insn->op == OP_ADD) || (insn->op == OP_MAD);
>
>    // add f32 LIMM cannot saturate
>    if (insn->op == OP_ADD && insn->sType == TYPE_F32) {
>       if (insn->getSrc(1)->asImm() &&
>           insn->getSrc(1)->reg.data.u32 & 0xfff)
>          return false;
>    }
>
> Note how it will say that sat is supported for SIMMs with FADD? So the
> compiler will generate those ops, but then the emitter won't be able
> to handle it.
>

Okay, I get it.