[Bug middle-end/71654] New: missing VRP optimization on c++ unsigned char and short expressions

[msebor at gcc dot gnu.org]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=71654

            Bug ID: 71654
           Summary: missing VRP optimization on c++ unsigned char and
                    short expressions
           Product: gcc
           Version: 7.0
            Status: UNCONFIRMED
          Severity: normal
          Priority: P3
         Component: middle-end
          Assignee: unassigned at gcc dot gnu.org
          Reporter: msebor at gcc dot gnu.org
  Target Milestone: ---

In C mode, GCC optimizes the both of the following functions into "return 0" as
one would expect.  But in C++ mode, it emits suboptimal code for guc: it does
the computation at runtime.  The same limitation exists if the type of the
argument is unsigned short.  From the dumps it looks as though the promotion
from the narrower type to int that's done in C++ before the inequality test
might be what prevents VRP from doing the expected thing.

$ cat vrp.c && /build/gcc-49905/gcc/xgcc -B /build/gcc-49905/gcc -O2 -S -Wall
-Wextra -fdump-tree-optimized=/dev/stdout -xc++ vrp.c
int gui (unsigned i)
{
  return i < 4 ? i & 4 : 0;
}

int guc (unsigned char i)
{
  return i < 4 ? i & 4 : 0;
}

;; Function int gui(unsigned int) (_Z3guij, funcdef_no=0, decl_uid=2248,
cgraph_uid=0, symbol_order=0)

int gui(unsigned int) (unsigned int i)
{
  <bb 2>:
  return 0;

}



;; Function int guc(unsigned char) (_Z3guch, funcdef_no=1, decl_uid=2251,
cgraph_uid=1, symbol_order=1)

Removing basic block 5
int guc(unsigned char) (unsigned char i)
{
  int _1;
  unsigned char _5;
  unsigned char _7;
  int prephitmp_8;

  <bb 2>:
  _1 = (int) i_3(D);
  if (_1 <= 3)
    goto <bb 3>;
  else
    goto <bb 4>;

  <bb 3>:
  _5 = i_3(D) & 4;

  <bb 4>:
  # _7 = PHI <_5(3), 0(2)>
  prephitmp_8 = (int) _7;
  return prephitmp_8;

}