[llvm-bugs] [Bug 47437] New: No vectorization - should use pmaddwd

[via llvm-bugs]

https://bugs.llvm.org/show_bug.cgi?id=47437

            Bug ID: 47437
           Summary: No vectorization - should use pmaddwd
           Product: libraries
           Version: trunk
          Hardware: PC
                OS: Linux
            Status: NEW
          Severity: enhancement
          Priority: P
         Component: Loop Optimizer
          Assignee: unassignedbugs at nondot.org
          Reporter: david.bolvansky at gmail.com
                CC: llvm-bugs at lists.llvm.org

typedef short int int16_t;
typedef int int32_t;

void
test_muladd (int32_t *       restrict d1,
             const int16_t * restrict s1,
             const int16_t * restrict s2,
             int n)
{
  int i;
  for (i = 0; i < n; i++) {
    d1[i] = ((s1[i*2 + 0] * s2[i*2 + 0])
             + (s1[i*2 + 1] * s2[i*2 + 1]));
  }
}

Clang -O3:
test_muladd:                            # @test_muladd
        test    ecx, ecx
        jle     .LBB0_7
        mov     r8d, ecx
        cmp     ecx, 1
        jne     .LBB0_3
        xor     eax, eax
        jmp     .LBB0_5
.LBB0_3:
        mov     r9d, r8d
        and     r9d, -2
        xor     eax, eax
.LBB0_4:                                # =>This Inner Loop Header: Depth=1
        movsx   r10d, word ptr [rsi + 4*rax]
        movsx   r11d, word ptr [rdx + 4*rax]
        imul    r11d, r10d
        movsx   r10d, word ptr [rsi + 4*rax + 2]
        movsx   ecx, word ptr [rdx + 4*rax + 2]
        imul    ecx, r10d
        add     ecx, r11d
        mov     dword ptr [rdi + 4*rax], ecx
        movsx   r10d, word ptr [rsi + 4*rax + 4]
        movsx   r11d, word ptr [rdx + 4*rax + 4]
        imul    r11d, r10d
        movsx   r10d, word ptr [rsi + 4*rax + 6]
        movsx   ecx, word ptr [rdx + 4*rax + 6]
        imul    ecx, r10d
        add     ecx, r11d
        mov     dword ptr [rdi + 4*rax + 4], ecx
        add     rax, 2
        cmp     r9, rax
        jne     .LBB0_4
.LBB0_5:
        test    r8b, 1
        je      .LBB0_7
        movsx   r8d, word ptr [rsi + 4*rax]
        movsx   ecx, word ptr [rdx + 4*rax]
        imul    ecx, r8d
        movsx   esi, word ptr [rsi + 4*rax + 2]
        movsx   edx, word ptr [rdx + 4*rax + 2]
        imul    edx, esi
        add     edx, ecx
        mov     dword ptr [rdi + 4*rax], edx
.LBB0_7:
        ret


ICC 19 produces:
..B1.16:                        # Preds ..B1.11 ..B1.16
        movdqu    xmm0, XMMWORD PTR [rsi+rdx*4]                 #12.15
        movdqu    xmm1, XMMWORD PTR [16+rsi+rdx*4]              #12.15
        pmaddwd   xmm0, XMMWORD PTR [r8+rdx*4]                  #12.29
        pmaddwd   xmm1, XMMWORD PTR [16+r8+rdx*4]               #12.29
        movdqu    XMMWORD PTR [rdi+rdx*4], xmm0                 #12.5
        movdqu    XMMWORD PTR [16+rdi+rdx*4], xmm1              #12.5
        add       rdx, 8                                        #11.3
        cmp       rdx, rcx                                      #11.3
        jb        ..B1.16 


Looking at -Rpass-missed logs, looks like cost model issue. 


Godbolt:
https://godbolt.org/z/E8Mf99

-- 
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/20200906/c9f79fa8/attachment-0001.html>