[llvm-bugs] [Bug 38280] New: Pointless loop unroll / vectorization

[via llvm-bugs]

https://bugs.llvm.org/show_bug.cgi?id=38280

            Bug ID: 38280
           Summary: Pointless loop unroll / vectorization
           Product: libraries
           Version: 6.0
          Hardware: PC
                OS: Windows NT
            Status: NEW
          Severity: enhancement
          Priority: P
         Component: Loop Optimizer
          Assignee: unassignedbugs at nondot.org
          Reporter: fabiang at radgametools.com
                CC: llvm-bugs at lists.llvm.org

Example C++ code for x86, simplified from a more complex use case:

// ---- begin

#include <stdint.h>
#include <stddef.h>
#include <emmintrin.h>

// neg_offs <= -8 required
void apply_delta(uint8_t *dst, const uint8_t *src, ptrdiff_t neg_offs, size_t
count)
{
    // Just provided for context
    while (count >= 8)
    {
        __m128i src_bytes = _mm_loadl_epi64((const __m128i *) src);
        __m128i pred_bytes = _mm_loadl_epi64((const __m128i *) (dst +
neg_offs));
        __m128i sum = _mm_add_epi8(src_bytes, pred_bytes);
        _mm_storel_epi64((__m128i *) dst, sum);

        dst += 8;
        src += 8;
        count -= 8;
    }

    // This is the loop in question
    while (count--)
    {
        *dst = *src + dst[neg_offs];
        dst++;
        src++;
    }
}

// ---- end

The bottom (tail) loop gets expanded into a giant monstrosity that attempts to
process 64 bytes at once, with various special-case paths for tail processing,
to handle cases where neg_offs > -64 (which means the obvious
64-elements-at-a-time loop would not work), etc.

The full code can be viewed at https://godbolt.org/g/yRThcs, I won't post it
here. :)

All of which is completely pointless because the tail loop will (as is easy to
see) only ever see count <= 7.

This is an extreme example, but I'm seeing this general pattern (a scalar tail
loop for a manually vectorized loop getting pointlessly auto-vectorized) a lot.

-- 
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/20180723/b7cd230c/attachment.html>