[clang++] operator==(vector of long, vector of long) -> vector of long long

[ImpleLee]

When comparing two variables of vector of long, clang++ returns a vector of long long, inconsistent with g++.
https://godbolt.org/z/db9zT6f4r

#include <type_traits>

using T [[gnu::vector_size(sizeof(long))]] = long;

static_assert(std::is_same_v<std::remove_reference_t<decltype((T{} == T{})[0])>, long>);

This makes the following code fails to compile with clang++ with the compilation options -march=x86-64-v4 -std=c++20 (previously reported as gcc bugzilla 118546).
Online link: https://godbolt.org/z/sxqsKrv9e .

#include <experimental/simd>
#include <cstdint>

using deduce_t_element = std::experimental::simd<
    std::uint64_t,
    std::experimental::simd_abi::deduce_t<std::uint64_t, 4>
>;

auto f(deduce_t_element e) {
    return e == 0;
}

[llvmbot]

@llvm/issue-subscribers-backend-x86

Author: Imple Lee (ImpleLee)

When comparing two variables of vector of long, clang++ returns a vector of long long, inconsistent with g++. https://godbolt.org/z/db9zT6f4r

#include <type_traits>

using T [[gnu::vector_size(sizeof(long))]] = long;

static_assert(std::is_same_v<std::remove_reference_t<decltype((T{} == T{})[0])>, long>);

This makes the following code fails to compile with clang++ with the compilation options -march=x86-64-v4 -std=c++20 (previously reported as gcc bugzilla 118546).
Online link: https://godbolt.org/z/sxqsKrv9e .

#include <experimental/simd>
#include <cstdint>

using deduce_t_element = std::experimental::simd<
    std::uint64_t,
    std::experimental::simd_abi::deduce_t<std::uint64_t, 4>
>;

auto f(deduce_t_element e) {
    return e == 0;
}

[llvmbot]

@llvm/issue-subscribers-clang-frontend

Author: Imple Lee (ImpleLee)

When comparing two variables of vector of long, clang++ returns a vector of long long, inconsistent with g++. https://godbolt.org/z/db9zT6f4r

#include <type_traits>

using T [[gnu::vector_size(sizeof(long))]] = long;

static_assert(std::is_same_v<std::remove_reference_t<decltype((T{} == T{})[0])>, long>);

This makes the following code fails to compile with clang++ with the compilation options -march=x86-64-v4 -std=c++20 (previously reported as gcc bugzilla 118546).
Online link: https://godbolt.org/z/sxqsKrv9e .

#include <experimental/simd>
#include <cstdint>

using deduce_t_element = std::experimental::simd<
    std::uint64_t,
    std::experimental::simd_abi::deduce_t<std::uint64_t, 4>
>;

auto f(deduce_t_element e) {
    return e == 0;
}

[shafik]

CC @mizvekov

[mizvekov]

The resulting type of the comparison is the signed equivalent of the original vector type.

But the way clang computes this is strange, it gets the size of the element type, and returns a signed vector of the same size,
by searching a list ordered by "largest types" first, so it looks for "long long" before "long".

On triples where sizeof(long long) == sizeof(long), it will always prefer long long, which is what you are seeing here.

[ImpleLee]

But the way clang computes this is strange, it gets the size of the element type, and returns a signed vector of the same size, by searching a list ordered by "largest types" first, so it looks for "long long" before "long".

Can we make it a type-to-type mapping, so that long always maps to long, and long long always maps to long long, etc? Usable element types in the vector type is limited, after all.

[ImpleLee]

The resulting type of the comparison is the signed equivalent of the original vector type.

But the way clang computes this is strange, it gets the size of the element type, and returns a signed vector of the same size, by searching a list ordered by "largest types" first, so it looks for "long long" before "long".

On triples where sizeof(long long) == sizeof(long), it will always prefer long long, which is what you are seeing here.

Hello, I would like to help to solve this issue, but I am not familiar with LLVM. Could you please give me any guide on where/how to start?

[mizvekov]

Can we make it a type-to-type mapping, so that long always maps to long, and long long always maps to long long, etc? Usable element types in the vector type is limited, after all.

Yeah, you can make sure this maps to the corresponding equivalent signed size, instead of just any signed type which is the same size.

Hello, I would like to help to solve this issue, but I am not familiar with LLVM. Could you please give me any guide on where/how to start?

Yeah, you can start by fixing this function

llvm-project/clang/lib/Sema/SemaExpr.cpp

Line 12882 in 1dfdd1e

QualType Sema::GetSignedVectorType(QualType V) {

I hope it is more or less self-evident what needs to be changed here. In any case, don't hesitate to ask for help.