missed optimization to simple vptest when using libstdc++ std::experimental::simd to detect all-zero pattern on x86-64 avx

[ImpleLee]

The following code uses libstdc++ experimental simd, and wants to detect several all-zero patterns that can be easily done with the vptest instructions. All the code is available at https://godbolt.org/z/Kx68E1T6v .

#include <experimental/simd>
#include <cstdint>
namespace stdx = std::experimental;

template <class T, std::size_t N>
using simd_of = stdx::simd<T, stdx::simd_abi::deduce_t<T, N>>;

using data_t = simd_of<std::int32_t, 4>;

bool simple_ptest(data_t x) {
    return all_of(x == 0);
}

bool ptest_and(data_t a, data_t b) {
    return all_of((a & b) == 0);
}

bool ptest_andn(data_t a, data_t b) {
    return all_of((a & ~b) == 0);
}

Equivalent assembly (hand-written):

simple_ptest:
        vptest  %xmm0, %xmm0
        sete    %al
        ret
ptest_and:
        vptest  %xmm0, %xmm1
        sete    %al
        ret
ptest_andn:
        vptest  %xmm0, %xmm1
        setc    %al
        ret

But clang++ generates the following code at -O3 -march=x86-64-v3.

simple_ptest(std::experimental::parallelism_v2::simd<int, std::experimental::parallelism_v2::simd_abi::_VecBuiltin<16>>):
        vpxor   xmm1, xmm1, xmm1
        vpcmpeqd        xmm0, xmm0, xmm1
        vpcmpeqd        xmm1, xmm1, xmm1
        vptest  xmm0, xmm1
        setb    al
        ret

ptest_and(std::experimental::parallelism_v2::simd<int, std::experimental::parallelism_v2::simd_abi::_VecBuiltin<16>>, std::experimental::parallelism_v2::simd<int, std::experimental::parallelism_v2::simd_abi::_VecBuiltin<16>>):
        vpand   xmm0, xmm1, xmm0
        vpxor   xmm1, xmm1, xmm1
        vpcmpeqd        xmm0, xmm0, xmm1
        vpcmpeqd        xmm1, xmm1, xmm1
        vptest  xmm0, xmm1
        setb    al
        ret

ptest_andn(std::experimental::parallelism_v2::simd<int, std::experimental::parallelism_v2::simd_abi::_VecBuiltin<16>>, std::experimental::parallelism_v2::simd<int, std::experimental::parallelism_v2::simd_abi::_VecBuiltin<16>>):
        vpandn  xmm0, xmm1, xmm0
        vpxor   xmm1, xmm1, xmm1
        vpcmpeqd        xmm0, xmm0, xmm1
        vpcmpeqd        xmm1, xmm1, xmm1
        vptest  xmm0, xmm1
        setb    al
        ret

reference: the same issue at gcc bugzilla (identified as a duplicate to another missed optimization).

[llvmbot]

@llvm/issue-subscribers-backend-x86

Author: Imple Lee (ImpleLee)

The following code uses libstdc++ experimental simd, and wants to detect several all-zero patterns that can be easily done with the vptest instructions. All the code is available at https://godbolt.org/z/Kx68E1T6v .

#include <experimental/simd>
#include <cstdint>
namespace stdx = std::experimental;

template <class T, std::size_t N>
using simd_of = stdx::simd<T, stdx::simd_abi::deduce_t<T, N>>;

using data_t = simd_of<std::int32_t, 4>;

bool simple_ptest(data_t x) {
    return all_of(x == 0);
}

bool ptest_and(data_t a, data_t b) {
    return all_of((a & b) == 0);
}

bool ptest_andn(data_t a, data_t b) {
    return all_of((a & ~b) == 0);
}

Equivalent assembly (hand-written):

simple_ptest:
        vptest  %xmm0, %xmm0
        sete    %al
        ret
ptest_and:
        vptest  %xmm0, %xmm1
        sete    %al
        ret
ptest_andn:
        vptest  %xmm0, %xmm1
        setc    %al
        ret

But clang++ generates the following code at -O3 -march=x86-64-v3.

simple_ptest(std::experimental::parallelism_v2::simd<int, std::experimental::parallelism_v2::simd_abi::_VecBuiltin<16>>):
        vpxor   xmm1, xmm1, xmm1
        vpcmpeqd        xmm0, xmm0, xmm1
        vpcmpeqd        xmm1, xmm1, xmm1
        vptest  xmm0, xmm1
        setb    al
        ret

ptest_and(std::experimental::parallelism_v2::simd<int, std::experimental::parallelism_v2::simd_abi::_VecBuiltin<16>>, std::experimental::parallelism_v2::simd<int, std::experimental::parallelism_v2::simd_abi::_VecBuiltin<16>>):
        vpand   xmm0, xmm1, xmm0
        vpxor   xmm1, xmm1, xmm1
        vpcmpeqd        xmm0, xmm0, xmm1
        vpcmpeqd        xmm1, xmm1, xmm1
        vptest  xmm0, xmm1
        setb    al
        ret

ptest_andn(std::experimental::parallelism_v2::simd<int, std::experimental::parallelism_v2::simd_abi::_VecBuiltin<16>>, std::experimental::parallelism_v2::simd<int, std::experimental::parallelism_v2::simd_abi::_VecBuiltin<16>>):
        vpandn  xmm0, xmm1, xmm0
        vpxor   xmm1, xmm1, xmm1
        vpcmpeqd        xmm0, xmm0, xmm1
        vpcmpeqd        xmm1, xmm1, xmm1
        vptest  xmm0, xmm1
        setb    al
        ret

reference: the same issue at gcc bugzilla (identified as a duplicate to another missed optimization).

[RKSimon]

Should be relatively trivial to add any missing patterns to combinePTESTCC