[Bug ipa/96482] [10/11 Regression] Combination of -finline-small-functions and ipa-cp optimisations causes incorrect values being passed to a function since r279523

[marxin at gcc dot gnu.org]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=96482

--- Comment #14 from Martin Liška <marxin at gcc dot gnu.org> ---
So in ltrans we end up with:

addr_to_index (struct nir_builder * b, struct nir_ssa_def * addr,
nir_address_format addr_format)
{
  unsigned int num_channels;
  unsigned int swizzle[16];
  unsigned int i;
  struct nir_ssa_def * D.7072;
  unsigned int c;
  struct nir_ssa_def * _9;

  <bb 2> [local count: 1073741824]:
  if (addr_format_3(D) == 3)
    goto <bb 3>; [20.24%]
  else
    goto <bb 4>; [79.76%]

  <bb 3> [local count: 1073741824]:
  c = 0;
  _9 = nir_swizzle (b_5(D), addr_6(D), &c, 1);
  c ={v} {CLOBBER};
  return _9;

  <bb 4> [count: 0]:
  __builtin_unreachable ();
}

that's fine, in WPA we identified that possible values are 3 and 5.
However ./src/gallium/targets/dri/libgallium_dri.so.ltrans65.ltrans.129t.ccp3
optimizes that to:

Folding statement: if (addr_format_3(D) == 3)
which is likely CONSTANT
Folding predicate addr_format_3(D) == 3 to 0
Folded into: if (0 != 0)

which is wrong.

The following code is executed:

          if (flag_tree_bit_ccp)
            {
              wide_int nonzero_bits = get_nonzero_bits (var);
              tree value;
              widest_int mask;

              if (SSA_NAME_VAR (var)
                  && TREE_CODE (SSA_NAME_VAR (var)) == PARM_DECL
                  && ipcp_get_parm_bits (SSA_NAME_VAR (var), &value, &mask))
                {
                  val.lattice_val = CONSTANT;
                  val.value = value;
                  val.mask = mask;
                  if (nonzero_bits != -1)
                    val.mask &= extend_mask (nonzero_bits,
                                             TYPE_SIGN (TREE_TYPE (var)));
                }

where range info for var is:

(gdb) p *ri
$2 = {
  ints = {
    m_precision = 32,
    m_max_len = 1 '\001',
    m_len = "\001\001\001",
    m_val = {3}
  }
}

(gdb) p nonzero_bits.dump()
[0x3], precision = 32

(gdb) p mask.dump()
[...,0x6], precision = 192

(gdb) p debug_tree(value)
 <integer_cst 0x7ffff7268e10 type <enumeral_type 0x7ffff720e738
nir_address_format> constant 5>

So that's we saw in WPA. So far so good, but now these 2 pieces of information
are combined:
                    val.mask &= extend_mask (nonzero_bits,
                                             TYPE_SIGN (TREE_TYPE (var)));

(gdb) p val.mask.dump()
[...,0x2], precision = 192

and folding happens :/