[Bug tree-optimization/91579] New: tailr1 pass creates redundant phi nodes

[jamborm at gcc dot gnu.org]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=91579

            Bug ID: 91579
           Summary: tailr1 pass creates redundant phi nodes
           Product: gcc
           Version: 10.0
            Status: UNCONFIRMED
          Severity: normal
          Priority: P3
         Component: tree-optimization
          Assignee: unassigned at gcc dot gnu.org
          Reporter: jamborm at gcc dot gnu.org
  Target Milestone: ---
              Host: x86_64-linux
            Target: x86_64-linux

When compiling the following testcase (at least at -O2 and higher),
the early tail-call pass creates redundant PHI nodes.  The problem
with them is that ipa-prop then cannot see through them and does not
discover that an unchanged scalar argument is passed to another call.
As a consequence we lose information in jump functions and may not
clone the function when we should.

typedef long unsigned int size_t;
typedef int (*compare_t)(const void *, const void *);
int partition (void *base, size_t nmemb, size_t size, compare_t cmp);
void
my_qsort (void *base, size_t nmemb, size_t size, compare_t cmp)
{
  int pt;
  if (nmemb > 1)
    {
      pt = partition (base, nmemb, size, cmp);
      my_qsort (base, pt + 1, size, cmp);
      my_qsort ((void*)((char*) base + (pt + 1) * size),
                nmemb - pt - 1, size, cmp);
    }
}

Results into:
  <bb 2> :
  # base_13 = PHI <base_6(D)(0), _9(3)>
  # nmemb_11 = PHI <nmemb_7(D)(0), _5(3)>
  # size_14 = PHI <size_20(D)(0), size_14(3)>
  # cmp_15 = PHI <cmp_21(D)(0), cmp_15(3)>

where the last two PHI nodes are superfluous.

(I think I have a patch to address it, let me test it.)