[cfe-commits] [PATCH] Add ptx32 and ptx64 target triples, as well as PTX builtins

[Eli Friedman]

On Wed, Apr 20, 2011 at 10:58 AM, Justin Holewinski
<justin.holewinski at gmail.com> wrote:
>
>
> On Wed, Apr 20, 2011 at 1:52 PM, Eli Friedman <eli.friedman at gmail.com>
> wrote:
>>
>> On Wed, Apr 20, 2011 at 10:50 AM, Eli Friedman <eli.friedman at gmail.com>
>> wrote:
>> > On Wed, Apr 20, 2011 at 10:14 AM, Justin Holewinski
>> > <justin.holewinski at gmail.com> wrote:
>> >> The attached patch adds target triples for the PTX back-end, and adds
>> >> the
>> >> currently implemented PTX intrinsics as builtin functions.  I would
>> >> like for
>> >> someone familiar with the Clang targets/triples and builtins code to
>> >> review
>> >> this patch.  I have been involved in the PTX back-end for several
>> >> months
>> >> now, but this is my first patch to Clang.
>> >
>> > +    PTXTargetInfo(const std::string& triple) : TargetInfo(triple) {
>> > +      TLSSupported = false;
>> > +      IntWidth = IntAlign = 32;
>> > +      LongWidth = LongLongWidth = LongAlign = LongLongAlign = 64;
>> > +    }
>> >
>> > You can skip setting IntWidth/IntAlign/LongLongWidth/LongLongAlign;
>> > the defaults are correct.
>
> Okay, will do.
>
>>
>> >
>> > +  class PTX32TargetInfo : public PTXTargetInfo {
>> > +  public:
>> > +  PTX32TargetInfo(const std::string& triple) : PTXTargetInfo(triple) {
>> > +      PointerWidth = PointerAlign = 32;
>> > +      DescriptionString
>> > +        = "e-p:32:32-i64:32:32-f64:32:32-v128:32:128-v64:32:64-n32:64";
>> > +    }
>> > +  };
>> >
>> > The target description string isn't using the same alignment for long
>> > and for double as the clang TargetInfo.  Also, the alignment for v64
>> > and v128 looks wrong.
>
> Would there be any issue with just removing the v64 and v128 specifiers for
> now?  The back-end does not currently support them.

That's fine.

>>
>> >
>> > BUILTIN(__builtin_ptx_read_tid_x, "i", "nc")
>> >
>> > The "c" means that the value returned by these builtins never changes;
>> > that doesn't seem right.
>> >
>
> This is an interesting case.  I assumed the "const" meant that successive
> calls to the builtin will return the same value (which is true).  However,
> different threads will see different values.  What is the recommended
> behavior here?  For a given piece of C code that will compile down to PTX,
> it is perfectly valid for constant propagation to eliminate any successive
> calls.

It should be fine for __builtin_ptx_read_tid_x, but it seems a bit
weird for  __builtin_ptx_read_clock, and very wrong for
__builtin_ptx_bar_sync.

>>
>> Oh, one more thing: this description unconditionally defines size_t
>> and friends to a 64-bit type; that isn't wrong, exactly, but probably
>> not what you want.
>
> Where exactly is this defined?  I'm not too familiar with Clang's TargetInfo
> class yet.

SizeType, PtrDiffType, IntPtrType.

-Eli