[Bug middle-end/61225] [5/6/7 Regression] Several new failures after r210458 on x86_64-*-* with -m32

[ubizjak at gmail dot com]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=61225

Uroš Bizjak <ubizjak at gmail dot com> changed:

           What    |Removed                     |Added
----------------------------------------------------------------------------
                 CC|                            |rguenth at gcc dot gnu.org

--- Comment #31 from Uroš Bizjak <ubizjak at gmail dot com> ---
What I find interesting here is, that following testcase:

--cut here--
void foo (void *);

#define F(T, OP, OPN) \
T *                     \
f##T##OPN (T *x, T y)   \
{                       \
  *x OP y;              \
  if (!*x)              \
    foo (x);            \
  return x;             \
}

#define G(T) \
F (T, |=, or)           \
F (T, ^=, xor)

G (int)
--cut here--

in fact also fails on x86_64 (-Os, -O2):

fintxor:
.LFB1:
        movl    (%rdi), %eax
        pushq   %rbx
.LCFI2:
        movq    %rdi, %rbx
        movl    %eax, %edx
        xorl    %esi, %edx
        cmpl    %esi, %eax
        movl    %edx, (%rdi)
        jne     .L6
        call    foo
.L6:
        movq    %rbx, %rax
        popq    %rbx
.LCFI3:
        ret

The failure is not detected, since the test scans only for TEST insn, not also
for CMP.

The difference between these two functions is already in the .optimized tree
dump between fintor (good):

fintor (int * x, int y)
{
  int _1;
  int _2;

  <bb 2> [100.00%]:
  _1 = *x_5(D);
  _2 = _1 | y_6(D);
  *x_5(D) = _2;
  if (_2 == 0)
    goto <bb 3>; [36.64%]
  else
    goto <bb 4>; [63.36%]

  <bb 3> [36.64%]:
  foo (x_5(D));

  <bb 4> [100.00%]:
  return x_5(D);

}

and fintxor (bad):

fintxor (int * x, int y)
{
  int _1;
  int _2;

  <bb 2> [100.00%]:
  _1 = *x_5(D);
  _2 = _1 ^ y_6(D);
  *x_5(D) = _2;
  if (_1 == y_6(D))
    goto <bb 3>; [36.64%]
  else
    goto <bb 4>; [63.36%]

  <bb 3> [36.64%]:
  foo (x_5(D));

  <bb 4> [100.00%]:
  return x_5(D);

}

Consequently, fintxor function is not expanded with memory operand, and things
remain unoptimized from the expand onward.