https://gcc.gnu.org/bugzilla/show_bug.cgi?id=115976

            Bug ID: 115976
           Summary: Missing uadd_sat/usub_sat vectorization
           Product: gcc
           Version: 15.0
            Status: UNCONFIRMED
          Keywords: missed-optimization
          Severity: normal
          Priority: P3
         Component: tree-optimization
          Assignee: unassigned at gcc dot gnu.org
          Reporter: pinskia at gcc dot gnu.org
  Target Milestone: ---
            Target: riscv

Take:
```
void f1(unsigned *__restrict__  a, unsigned * __restrict__ b)
{
        for(int i = 0;i < 1024;i ++)
        {
          unsigned long long ta = a[i];
          unsigned long long tb = b[i];
          unsigned long long tt = ta + tb;
          if (tt > -1u)
            tt = -1u;

          a[i] = tt;
        }
}
void f2(unsigned *__restrict__  a, unsigned * __restrict__ b)
{
        for(int i = 0;i < 1024;i ++)
        {
          unsigned long long ta = a[i];
          unsigned long long tb = b[i];
          unsigned long long tt = ta - tb;
          if (tt < 0)
            tt = 0;
          if (tt > -1u)
            tt = -1u;

          a[i] = tt;
        }
}
```

f1 should produce a SAT_ADD (unsigned) but currently it produces an widening
add followed by a SAT_TRUNC.

f2 should produce a SAT_SUB (unsigned) but currently it produces an widening
sub followed by a SAT_TRUNC.

Note this is forwarded from https://github.com/llvm/llvm-project/pull/99250.

Reply via email to