[SLP] gcc generate better code than clang base on stride offset

[vfdff]

• test case, see https://gcc.godbolt.org/z/9E5beqnWW

void foo ( float *restrict  fi, real *restrict  f, int ci) {
  
    /* Add accumulated i-forces to the force array */
    for (int i = 0; i < UNROLLI; i++) {
         for (int d = 0; d < DIM; d++) {
             f[(ci*UNROLLI+i)*F_STRIDE+d] += fi[i*FI_STRIDE+d];
         }
    }

  return;
}

• gcc: SLP
• llvm: unroll

[vfdff]

• It is possible that optimizations related to associative on llvm need to be enhanced, by comparing foo_1 and foo_2,
  see detail https://gcc.godbolt.org/z/jdMcr8ozx

void foo_1 ( float *restrict  fi, real *restrict  f, int ci) {
    f[(ci*4+0)*3+0] += fi[0];
    f[(ci*4+0)*3+1] += fi[1];
    f[(ci*4+0)*3+2] += fi[2];

    f[(ci*4+1)*3+0] += fi[3];
    f[(ci*4+1)*3+1] += fi[4];
    f[(ci*4+1)*3+2] += fi[5];

    f[(ci*4+2)*3+0] += fi[6];
    f[(ci*4+2)*3+1] += fi[7];
    f[(ci*4+2)*3+2] += fi[8];

    f[(ci*4+3)*3+0] += fi[9];
    f[(ci*4+3)*3+1] += fi[10];
    f[(ci*4+3)*3+2] += fi[11];
  return;
}

void foo_2 ( float *restrict  fi, real *restrict  f, int ci) {
    f[ci*12+0] += fi[0];
    f[ci*12+1] += fi[1];
    f[ci*12+2] += fi[2];

    f[ci*12+3] += fi[3];
    f[ci*12+4] += fi[4];
    f[ci*12+5] += fi[5];

    f[ci*12+6] += fi[6];
    f[ci*12+7] += fi[7];
    f[ci*12+8] += fi[8];

    f[ci*12+9] += fi[9];
    f[ci*12+10] += fi[10];
    f[ci*12+11] += fi[11];
  return;
}

[vfdff]

Add https://reviews.llvm.org/D132658 for the above function foo_1

[vfdff]

should still open as the origal foo is not fixed, D132658 only fixed the foo_1

[vfdff]

• Now the origal foo is also SLP optimized.

• But now the llvm assemble of foo_1 seems much more
  record the llvm's assemble to compare the future optimized assemble.

foo_1:                                  // @foo_1
        add     w8, w2, w2, lsl #1
        ldr     q3, [x0, #32]
        lsl     w8, w8, #2
        sbfiz   x9, x8, #2, #32
        add     w10, w8, #4
        ldp     q1, q2, [x0]
        add     w8, w8, #8
        sbfiz   x10, x10, #2, #32
        sbfiz   x8, x8, #2, #32
        ldr     q0, [x1, x9]
        fadd    v0.4s, v0.4s, v1.4s
        str     q0, [x1, x9]
        ldr     q0, [x1, x10]
        ldr     q1, [x1, x8]
        fadd    v0.4s, v0.4s, v2.4s
        fadd    v1.4s, v1.4s, v3.4s
        str     q0, [x1, x10]
        str     q1, [x1, x8]
        ret

[vfdff]

• For foo_1, llvm get much better performance with -mllvm -aarch64-enable-gep-opt=true, https://gcc.godbolt.org/z/GfocnG45W, and it is changed since commit 201b785

foo_1:                                  // @foo_1
        add     w8, w2, w2, lsl #1
        ldr     q2, [x0, #32]
        lsl     w8, w8, #2
        ldp     q0, q1, [x0]
        add     x8, x1, w8, sxtw #2
        ldp     q3, q4, [x8]
        fadd    v0.4s, v3.4s, v0.4s
        ldr     q5, [x8, #32]
        fadd    v1.4s, v4.4s, v1.4s
        fadd    v2.4s, v5.4s, v2.4s
        stp     q0, q1, [x8]
        str     q2, [x8, #32]
        ret