Should simd_mask support shift operators

[danieltowner]

A customer has requested that simd_mask should be shiftable, both in binary and compound assignment forms. Any objection to adding those?

[DenisYaroshevskiy]

I'm not 100% what that means.

There are a by lane operations and there is what we call slide_left, slide_right - which is a swizzle (thnk bslli)

Which one is it?

[mattkretz]

There's a clear answer:

T x;
auto y = (x == 0) << 1;

The type of y determines the behavior of simd. y is going to be an int, because bool promotes to int before the shift operator is called.
So

friend constexpr simd_type
operator<<(simd_mask, int);

This is also consistent with the other operator overloads in simd_mask.

No objection from my side. I'd be interested in seeing a use case, though. :-)

[DenisYaroshevskiy]

The only reason it works for bool is because of an implicit conversion. I think that in general implicit conversions for simd is an awful idea.

But here - is there a mask implicit conversion to int?

[mattkretz]

See Section 4.8 of P1928R3. This was subsequently implemented in P1928R4.

[DenisYaroshevskiy]

template <class U, class A>
constexpr explicit(sizeof(U) != sizeof(T)) operator simd<U, A>() const noexcept;

?
I don't understand this one.

If implicit conversions from bool to int are to stay, I don't mind this. I didn't know about bool shift.

[danieltowner]

I'm not 100% what that means.

There are a by lane operations and there is what we call slide_left, slide_right - which is a swizzle (thnk bslli)

Which one is it?

Sorry, that was bad of me. I mean the second. What you call slide_left, or like https://en.cppreference.com/w/cpp/algorithm/shift.

std::bitset has these as well (https://en.cppreference.com/w/cpp/utility/bitset/operator_ltltgtgt) and I think in many user's minds (mine included) simd_mask and std::bitset are very comparable, so if you can shift bitsets, then you should be able to similarly shift a simd_mask. This is probably compounded by AVX-512 programmers who also know that predicate masks on that target can be shifted too (e.g., _kshiftri_mask16).

[DenisYaroshevskiy]

O_o. That is a very useful operation but honestly I'd expect it to be called the same way as it would for simd.

• Shift like tjis would break operators work by lane

Also - compile time shift only I'd say, otherwise even on avx2 this sucks.

[mattkretz]

@danieltowner Why isn't permute(k, shift_by<2>) acceptable here?

[danieltowner]

It probably could be acceptable, it's just not as obvious perhaps, especially for someone coming from a target with bit masks. At the moment the code generated doesn't do very well either, but I guess the compiler could be taught to do better. The best code is generated by

The best code is generated by assuming that https://isocpp.org/files/papers/P2876R0.html#compact_ctor is available (simd_mask(original.to_bitint() << 5). Then it maps to a real shift of the bitmask value itself on targets that support it.

If we had named functions, then shift_left and shift_right could be used to do the operation on a mask too.