[Bug target/79709] New: Subobtimal code with -mavx and explicit vector

[tkoenig at gcc dot gnu.org]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=79709

            Bug ID: 79709
           Summary: Subobtimal code with -mavx and explicit vector
           Product: gcc
           Version: 7.0.1
            Status: UNCONFIRMED
          Severity: normal
          Priority: P3
         Component: target
          Assignee: unassigned at gcc dot gnu.org
          Reporter: tkoenig at gcc dot gnu.org
  Target Milestone: ---

For the following code

typedef double v4do __attribute__((vector_size (32)));
typedef long int v4i __attribute__((vector_size (32)));

#define VSET(vect,val) do { vect[0]=val; vect[1]=val; vect[2]=val; vect[3]=val;
} while (0)
void foo(v4do cx, v4do cy, v4i *r)
{
  v4do x, y, xn, yn;
  v4i add, res;
  v4do two, four;
  long int done;

  VSET(res, 0L);
  VSET(two, 2.0);
  VSET(four, 4.0);
  x = cx;
  y = cy;
  done = 0;
  while (1)
    {
      xn = x*x - y*y + cx;
      yn = two*x*y + cy;
      add = xn+xn + yn*yn < four;
      res += add;
      if (add[0] == 0 || add[1] == 0 || add[2] || add[3])
        break;
      x = xn;
      y = yn;
    }
  *r = res;
}

gcc compares strange code.  The loop is translated with 7.0.1 20170212
with "gcc -O3 -S -mavx v.c" into

.L14:
        vpextrq $1, %xmm2, %rax
        testq   %rax, %rax
        je      .L2
        vmovdqa -48(%rbp), %ymm5
        vextractf128    $0x1, %ymm5, %xmm2
        vmovq   %xmm2, %rax
        testq   %rax, %rax
        jne     .L2
        vpextrq $1, %xmm2, %rax
        vmovapd %ymm3, %ymm5
        testq   %rax, %rax
        jne     .L2
.L3:
        vmulpd  %ymm5, %ymm5, %ymm3
        vmulpd  %ymm8, %ymm5, %ymm5
        vsubpd  %ymm6, %ymm3, %ymm3
        vmulpd  %ymm4, %ymm5, %ymm4
        vaddpd  %ymm0, %ymm3, %ymm3
        vaddpd  %ymm1, %ymm4, %ymm4
        vaddpd  %ymm3, %ymm3, %ymm2
        vmulpd  %ymm4, %ymm4, %ymm6
        vaddpd  %ymm6, %ymm2, %ymm2
        vcmpltpd        %ymm7, %ymm2, %ymm5
        vmovapd %ymm5, -48(%rbp)
        vmovdqa -48(%rbp), %xmm5
        vpaddq  -112(%rbp), %xmm5, %xmm5
        vmovaps %xmm5, -80(%rbp)
        vmovdqa -32(%rbp), %xmm5
        vpaddq  -96(%rbp), %xmm5, %xmm2
        vmovaps %xmm2, -64(%rbp)
        vmovdqa -80(%rbp), %ymm2
        vmovdqa %ymm2, -112(%rbp)
        vmovdqa -48(%rbp), %xmm2
        vmovq   %xmm2, %rax
        testq   %rax, %rax
        jne     .L14

which contains quite a few unnecessary instructions for moving stuff around.

By comparision, clang translates the inner loop to

.LBB0_1:                                # =>This Inner Loop Header: Depth=1
        vmulpd  %ymm5, %ymm5, %ymm6
        vmulpd  %ymm4, %ymm4, %ymm7
        vsubpd  %ymm7, %ymm6, %ymm6
        vaddpd  %ymm5, %ymm5, %ymm7
        vaddpd  %ymm0, %ymm6, %ymm5
        vmulpd  %ymm7, %ymm4, %ymm4
        vaddpd  %ymm1, %ymm4, %ymm4
        vaddpd  %ymm5, %ymm5, %ymm6
        vmulpd  %ymm4, %ymm4, %ymm7
        vaddpd  %ymm7, %ymm6, %ymm6
        vcmpltpd        %ymm8, %ymm6, %ymm6
        vextractf128    $1, %ymm6, %xmm7
        vextractf128    $1, %ymm2, %xmm3
        vpaddq  %xmm3, %xmm7, %xmm3
        vpaddq  %xmm2, %xmm6, %xmm2
        vinsertf128     $1, %xmm3, %ymm2, %ymm2
        vmovq   %xmm7, %rax
        vpextrq $1, %xmm7, %rcx
        orq     %rax, %rcx
        jne     .LBB0_4
# BB#2:                                 #   in Loop: Header=BB0_1 Depth=1
        vpextrq $1, %xmm6, %rax
        testq   %rax, %rax
        je      .LBB0_4
# BB#3:                                 #   in Loop: Header=BB0_1 Depth=1
        vmovq   %xmm6, %rax
        testq   %rax, %rax
        jne     .LBB0_1

which looks much more straighforward, and should be faster.