[Bug tree-optimization/66623] New: Unsafe FP math reduction used in strict math mode

[david.sherwood at arm dot com]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=66623

            Bug ID: 66623
           Summary: Unsafe FP math reduction used in strict math mode
           Product: gcc
           Version: 6.0
            Status: UNCONFIRMED
          Severity: normal
          Priority: P3
         Component: tree-optimization
          Assignee: unassigned at gcc dot gnu.org
          Reporter: david.sherwood at arm dot com
  Target Milestone: ---

Created attachment 35825
  --> https://gcc.gnu.org/bugzilla/attachment.cgi?id=35825&action=edit
Unsafe FP math reduction example

I've found a bug with reductions for Neon whereby we change the ordering
of FP computation in strict math mode. The example looks like this:

float foo (float *__restrict__ i)
{
  float l = 0;

  for (int a = 0; a < 4; a++)
    for (int b = 0; b < 4; b++)
      l += i[b];
  return l;
}

when compiled with the flags

-O2 -ftree-vectorize -fno-inline -march=armv8-a

we generate the asm:

        movi    v0.4s, 0
        mov     x1, x0
        mov     w0, 0
.L2:
        ldr     s1, [x1, w0, sxtw 2]
        add     w0, w0, 1
        cmp     w0, 4
        dup     v1.4s, v1.s[0]
        fadd    v0.4s, v0.4s, v1.4s
        bne     .L2
        faddp   v0.4s, v0.4s, v0.4s
        faddp   v0.4s, v0.4s, v0.4s

which is (i[0] + i[1] + ...) + (i[0] + i[1] + ...) + ... We know that in
general
"(a + b) + (a + b)" is not guaranteed to be the same as "((a + b) + a) + b".