[Bug rtl-optimization/66489] New: combine fails to merge insns if some are reused later on

[ktkachov at gcc dot gnu.org]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=66489

            Bug ID: 66489
           Summary: combine fails to merge insns if some are reused later
                    on
           Product: gcc
           Version: 6.0
            Status: UNCONFIRMED
          Keywords: missed-optimization
          Severity: normal
          Priority: P3
         Component: rtl-optimization
          Assignee: unassigned at gcc dot gnu.org
          Reporter: ktkachov at gcc dot gnu.org
  Target Milestone: ---

Consider code:
int f2(int a, int b, int c, int d)
{
  a = -a;
  b += a * c;
  c += a * d;
  return b ^ c;
}

On aarch64 this could be written with to msub instructions, RTL code:

(set (reg/i:SI 0 x0)
     (minus:SI
       (reg:SI 1 x1 [ b ])
       (mult:SI (reg:SI 0 x0 [ a ])
                (reg:SI 2 x2 [ c ])))) 

However, combine doesn't merge the neg and the multiply-adds and generates:
f2:
        neg     w4, w0
        madd    w0, w4, w2, w1
        madd    w3, w4, w3, w2
        eor     w0, w0, w3
        ret


If we modify the code to only do a single multiply-accumulate:
int f2(int a, int b, int c, int d)
{
  a = -a;
  b += a * c;
  return b;
}

Then they the expected single msub instruction is generated.
I think this is due to combine being scared of the negated 'a' being used in
two places.