[Mesa-dev] [PATCH v2 09/29] nir: Make boolean conversions sized just like the others

[Jason Ekstrand]

On December 13, 2018 04:07:56 Iago Toral <itoral at igalia.com> wrote:

> On Thu, 2018-12-06 at 13:45 -0600, Jason Ekstrand wrote:
> (...)
>> diff --git a/src/compiler/nir/nir_builder.h
>> b/src/compiler/nir/nir_builder.h
>> index 30fa1d7ec8b..e0cdcd4ba23 100644
>> --- a/src/compiler/nir/nir_builder.h
>> +++ b/src/compiler/nir/nir_builder.h
>> @@ -963,6 +963,18 @@ nir_load_param(nir_builder *build, uint32_t
>> param_idx)
>>
>> #include "nir_builder_opcodes.h"
>>
>> +static inline nir_ssa_def *
>> +nir_f2b(nir_builder *build, nir_ssa_def *f)
>> +{
>> +   return nir_f2b32(build, f);
>> +}
>> +
>> +static inline nir_ssa_def *
>> +nir_i2b(nir_builder *build, nir_ssa_def *i)
>> +{
>> +   return nir_i2b32(build, i);
>> +}
>> +
>
> any particular reason why wanted these instead of just calling the 32-
> bit opcode directly from the caller?

In my 1-bit booleans series, it lets me switch most of NIR without having 
to manually edit every single place we do a x2b conversion.

> I need to do b2f conversions in a couple of places where the
> destination can be 16, 32 or 64 and I think it is more convenient to
> have helpers that take the destination bit-size as parameter and then
> emit the appropriate opcode instead. What do you think?

For b2f, I totally agree.  In fact, I've considered adding a small bit of 
codegen to nir_builder to add such a pole of helpers. One of my patch 
series (I don't remember which at this point) adds i2i and u2u helpers that 
take a bit size like you suggest.


--Jason