[Bug target/79722] New: Missed opportunity for fused multiply/add with avx2

[tkoenig at gcc dot gnu.org]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=79722

            Bug ID: 79722
           Summary: Missed opportunity for fused multiply/add with avx2
           Product: gcc
           Version: unknown
            Status: UNCONFIRMED
          Severity: normal
          Priority: P3
         Component: target
          Assignee: unassigned at gcc dot gnu.org
          Reporter: tkoenig at gcc dot gnu.org
  Target Milestone: ---

Created attachment 40835
  --> https://gcc.gnu.org/bugzilla/attachment.cgi?id=40835&action=edit
Output of gcc -Ofast -mavx2 -S -o bar-gcc.s bar.c

The test case is the same as PR 79709:

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;
}

With gcc, the inner loop is tranlsated into

.L13:
        vpextrq $1, %xmm2, %rax
        testq   %rax, %rax
        je      .L2
        vextracti128    $0x1, %ymm2, %xmm2
        vmovq   %xmm2, %rax
        testq   %rax, %rax
        jne     .L2
        vpextrq $1, %xmm2, %rax
        vmovapd %ymm4, %ymm3
        testq   %rax, %rax
        jne     .L2
.L3:
        vmulpd  %ymm3, %ymm3, %ymm4
        vmulpd  %ymm9, %ymm3, %ymm3
        vaddpd  %ymm0, %ymm4, %ymm4
        vmulpd  %ymm7, %ymm3, %ymm3
        vsubpd  %ymm10, %ymm4, %ymm4
        vaddpd  %ymm1, %ymm3, %ymm9
        vaddpd  %ymm4, %ymm4, %ymm2
        vmulpd  %ymm9, %ymm9, %ymm10
        vaddpd  %ymm10, %ymm2, %ymm2
        vcmpltpd        %ymm6, %ymm2, %ymm2
        vmovq   %xmm2, %rax
        vpaddq  %ymm2, %ymm8, %ymm8
        testq   %rax, %rax
        jne     .L13

icc -O3 -march=core-avx2 -S results in

..B1.6:                         # Preds ..B1.5
                                # Execution count [6.94e-01]
        vmovdqa   %ymm11, %ymm5                                 #27.7
                                # LOE rbx rdi r12 r13 r14 r15 ymm0 ymm1 ymm2
ymm3 ymm4 ymm5 ymm6 ymm7
..B1.2:                         # Preds ..B1.6 ..B1.1
                                # Execution count [1.69e+00]
        vmovaps   %ymm4, %ymm11                                 #21.24
        vfmsub213pd %ymm6, %ymm4, %ymm11                        #21.24
        vfmsub231pd %ymm5, %ymm5, %ymm11                        #21.24
        vmulpd    %ymm3, %ymm5, %ymm5                           #22.16
        vaddpd    %ymm11, %ymm11, %ymm8                         #23.16
        vfmadd213pd %ymm7, %ymm5, %ymm4                         #22.22
        vfmadd231pd %ymm4, %ymm4, %ymm8                         #23.24
        vcmpltpd  %ymm2, %ymm8, %ymm9                           #23.29
        vandpd    %ymm9, %ymm1, %ymm10                          #23.29
        vmovups   %ymm10, -64(%rsp)                             #23.7
        vpaddq    %ymm10, %ymm0, %ymm0                          #24.7
        cmpq      $0, -64(%rsp)                                 #25.21
        je        ..B1.7        # Prob 20%                      #25.21
                                # LOE rbx rdi r12 r13 r14 r15 ymm0 ymm1 ymm2
ymm3 ymm4 ymm6 ymm7 ymm11
..B1.3:                         # Preds ..B1.2
                                # Execution count [1.36e+00]
        cmpq      $0, -56(%rsp)                                 #25.36
        je        ..B1.7        # Prob 20%                      #25.36
                                # LOE rbx rdi r12 r13 r14 r15 ymm0 ymm1 ymm2
ymm3 ymm4 ymm6 ymm7 ymm11
..B1.4:                         # Preds ..B1.3
                                # Execution count [1.08e+00]
        cmpq      $0, -48(%rsp)                                 #25.45
        jne       ..B1.7        # Prob 20%                      #25.45
                                # LOE rbx rdi r12 r13 r14 r15 ymm0 ymm1 ymm2
ymm3 ymm4 ymm6 ymm7 ymm11
..B1.5:                         # Preds ..B1.4
                                # Execution count [8.67e-01]
        cmpq      $0, -40(%rsp)                                 #25.55
        je        ..B1.6        # Prob 80%                      #25.55
                                # LOE rbx rdi r12 r13 r14 r15 ymm0 ymm1 ymm2
ymm3 ymm4 ymm6 ymm7 ymm11

where icc uses eight double precision floating point operations
vs. gcc's ten.