[llvm-bugs] [Bug 48662] New: SLP vectorizer produces lower performance code

[via llvm-bugs]

https://bugs.llvm.org/show_bug.cgi?id=48662

            Bug ID: 48662
           Summary: SLP vectorizer produces lower performance code
           Product: libraries
           Version: trunk
          Hardware: PC
                OS: Linux
            Status: NEW
          Severity: enhancement
          Priority: P
         Component: Scalar Optimizations
          Assignee: unassignedbugs at nondot.org
          Reporter: chfast at gmail.com
                CC: llvm-bugs at lists.llvm.org

The code just checks if array uint64_t[4] has all zeros.

using u64 = unsigned long;

struct u256
{
    u64 w[4];
};


bool iszero(const u256& u)
{
    return (u.w[0] == 0) & (u.w[1] == 0) & (u.w[2] == 0) & (u.w[3] == 0);
}


Without SLP vectorization (-fno-slp-vectorize) the compiler produces nice OR
folding:

iszero(u256 const&):                       # @iszero(u256 const&)
        mov     rax, qword ptr [rdi + 8]
        or      rax, qword ptr [rdi]
        or      rax, qword ptr [rdi + 16]
        or      rax, qword ptr [rdi + 24]
        sete    al
        ret

With SLP vectorization (default):

iszero(u256 const&):                       # @iszero(u256 const&)
        movdqu  xmm0, xmmword ptr [rdi]
        movdqu  xmm1, xmmword ptr [rdi + 16]
        por     xmm1, xmm0
        pxor    xmm0, xmm0
        pcmpeqb xmm0, xmm1
        pmovmskb        eax, xmm0
        cmp     eax, 65535
        sete    al
        ret


>From llvm-mca the vectorized code may have higher throughput but higher
latency. In my real life usage of similar code I can definitely notice the
performance degradation.

https://godbolt.org/z/z19EPc

-- 
You are receiving this mail because:
You are on the CC list for the bug.
-------------- next part --------------
An HTML attachment was scrubbed...
URL: <http://lists.llvm.org/pipermail/llvm-bugs/attachments/20210105/cc2fc49d/attachment.html>