[LLVMbugs] [Bug 21778] New: NEON intrinsics prevent removing redundant store and load on armv7

[bugzilla-daemon at llvm.org]

http://llvm.org/bugs/show_bug.cgi?id=21778

            Bug ID: 21778
           Summary: NEON intrinsics prevent removing redundant store and
                    load on armv7
           Product: libraries
           Version: trunk
          Hardware: Macintosh
                OS: All
            Status: NEW
          Severity: normal
          Priority: P
         Component: Scalar Optimizations
          Assignee: unassignedbugs at nondot.org
          Reporter: simontaylor1 at ntlworld.com
                CC: llvmbugs at cs.uiuc.edu
    Classification: Unclassified

I noticed that LLVM was unable to optimize redundant stores and loads when I
added some NEON intrinsics into my 4x4 matrix multiply function.

I've created a smaller example that shows the same problem. This is using
"Apple LLVM version 6.0 (clang-600.0.56) (based on LLVM 3.5svn)" (from the
latest XCode bundle with iOS SDK), targeting armv7, using -O3.

Here is the test code without intrinsics:

struct vec4
{
  float data[4];
};

vec4 operator* (const vec4& a, const vec4& b)
{
  vec4 result;
  for(int i = 0; i < 4; ++i)
    result.data[i] = a.data[i] * b.data[i];

  return result;
}

void TestVec4Multiply(vec4& a, vec4& b, vec4& result)
{
  result = a * b;
}

void TestVec4Multiply3(vec4& a, vec4& b, vec4& c, vec4& result)
{
  result = a * b * c;
}

In this case, the vectorizer actually generates the optimal code:

__Z16TestVec4MultiplyR4vec4S0_S0_:
@ BB#0:
    vld1.32    {d16, d17}, [r1]
    vld1.32    {d18, d19}, [r0]
    vmul.f32    q8, q9, q8
    vst1.32    {d16, d17}, [r2]
    bx    lr

__Z17TestVec4Multiply3R4vec4S0_S0_S0_:
@ BB#0:
    vld1.32    {d16, d17}, [r1]
    vld1.32    {d18, d19}, [r0]
    vmul.f32    q8, q9, q8
    vld1.32    {d18, d19}, [r2]
    vmul.f32    q8, q8, q9
    vst1.32    {d16, d17}, [r3]
    bx    lr

With my actual matrix multiply code the vectorizer is not as successful, hence
wanting to help out the compiler with some intrinsics. Here's a replacement of
the operator* with an implementation using NEON intrinsics:

vec4 operator* (const vec4& a, const vec4& b)
{
  vec4 result;

  float32x4_t result_data = vmulq_f32(vld1q_f32(a.data), vld1q_f32(b.data));
  vst1q_f32(result.data, result_data);

  return result;
}

Unfortunately the generated code now has some redundant stores and loads:

__Z16TestVec4MultiplyR4vec4S0_S0_:
@ BB#0:
    sub    sp, #16
    vld1.32    {d16, d17}, [r1]
    vld1.32    {d18, d19}, [r0]
    mov    r0, sp
    vmul.f32    q8, q9, q8
    vst1.32    {d16, d17}, [r0]
    vld1.32    {d16, d17}, [r0]
    vst1.32    {d16, d17}, [r2]
    add    sp, #16
    bx    lr

__Z17TestVec4Multiply3R4vec4S0_S0_S0_:
@ BB#0:
    sub    sp, #32
    vld1.32    {d16, d17}, [r1]
    vld1.32    {d18, d19}, [r0]
    mov    r0, sp
    vmul.f32    q8, q9, q8
    vst1.32    {d16, d17}, [r0]
    vld1.32    {d16, d17}, [r2]
    vld1.32    {d18, d19}, [r0]
    add    r0, sp, #16
    vmul.f32    q8, q9, q8
    vst1.32    {d16, d17}, [r0]
    vld1.32    {d16, d17}, [r0]
    vst1.32    {d16, d17}, [r3]
    add    sp, #32
    bx    lr

These seem to be especially bad news on many ARM cores. See here:
http://lists.freedesktop.org/archives/pixman/2011-August/001398.html

In my testing of 4x4 matrix multiply, the version with the temporaries ends up
about 3x slower than code that has them eliminated.

-- 
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/20141208/86f24ac1/attachment.html>