Re: [PATCH i386 AVX512] [81/n] Add new built-ins.

[Richard Biener <richard.guenther@gmail.com>]

On Mon, Oct 20, 2014 at 3:50 PM, Jakub Jelinek <jakub@redhat.com> wrote:
> On Mon, Oct 20, 2014 at 05:41:25PM +0400, Kirill Yukhin wrote:
>> Hello,
>> This patch adds (almost) all built-ins needed by
>> AVX-512VL,BW,DQ intrinsics.
>>
>> Main questionable hunk is:
>>
>> diff --git a/gcc/tree-core.h b/gcc/tree-core.h
>> index b69312b..a639487 100644
>> --- a/gcc/tree-core.h
>> +++ b/gcc/tree-core.h
>> @@ -1539,7 +1539,7 @@ struct GTY(()) tree_function_decl {
>>       DECL_FUNCTION_CODE.  Otherwise unused.
>>       ???  The bitfield needs to be able to hold all target function
>>         codes as well.  */
>> -  ENUM_BITFIELD(built_in_function) function_code : 11;
>> +  ENUM_BITFIELD(built_in_function) function_code : 12;
>>    ENUM_BITFIELD(built_in_class) built_in_class : 2;
>>
>>    unsigned static_ctor_flag : 1;
>
> Well, decl_with_vis has 15 unused bits, so instead of growing
> FUNCTION_DECL significantly, might be better to move one of the
> flags to decl_with_vis and just document that it applies to FUNCTION_DECLs
> only.  Or move some flag to cgraph if possible.
>
> But seeing e.g.
>        IX86_BUILTIN_FIXUPIMMPD256, IX86_BUILTIN_FIXUPIMMPD256_MASK,
>        IX86_BUILTIN_FIXUPIMMPD256_MASKZ
> etc. I wonder if you really need that many builtins, weren't we adding
> for avx512f just single builtin instead of 3 different ones, always
> providing mask argument and depending on whether it is all ones, etc.
> figuring out what kind of masking should be performed?

If only we had no lang-specific flags in tree_base we could use
the same place as we use for internal function code ...

But yes, not using that many builtins in the first place is preferred
for example by making them type-generic and/or variadic.

Richard.

>         Jakub