Re: [PATCH v3] enum_flags to_string (was: Re: [PATCH v2 07/29] Thread options & clone events (core + remote))

[Simon Marchi <simark@simark.ca>]

On 10/28/22 14:23, Pedro Alves wrote:
> On 2022-10-28 4:59 p.m., Simon Marchi wrote:
>> On 2022-10-28 07 h 08, Pedro Alves wrote:
>>> On 2022-10-28 11:26 a.m., Pedro Alves wrote:
> ...
>>> +  /* Check string conversion.  */
>>> +  {
>>> +    SELF_CHECK (to_string (0) == "0x0 []");
>>> +    SELF_CHECK (to_string (FLAG1) == "0x2 [FLAG1]");
>>> +    SELF_CHECK (to_string (FLAG1 | FLAG3) == "0xa [FLAG1 FLAG3]");
>>> +    SELF_CHECK (to_string (FLAG1 | FLAG2 | FLAG3) == "0xe [FLAG1 FLAG3 0x4]");
>>> +    SELF_CHECK (to_string (FLAG2) == "0x4 [0x4]");
>>> +    SELF_CHECK (to_string (test_flag (0xff)) == "0xff [FLAG1 FLAG3 0xf5]");
>>
>> Ish, this crashes with:
>>
>> /home/smarchi/src/binutils-gdb/gdb/../gdbsupport/enum-flags.h:191:12: runtime error: load of value 255, which is not a valid value for type 'test_flag'
> 
> Ah.  That's because I missed that test_flag doesn't have an explicit underlying
> type, as in "enum test_flag : underlying".  The rules are different if there's
> an explicit underlying type or not.  enums with an explicit type can hold any value
> of the underlying type.  Best to fix the testcase.  We could either remove
> that particular test, which is not really that important, as what it tests is
> exercised by the previous two tests, or switch to test_uflags instead, which
> does have an explicit underlying type.  I did the latter.

Ok, didn't know that.  This is fine with me.

> @@ -183,6 +194,52 @@ class enum_flags
>    /* Binary operations involving some unrelated type (which would be a
>       bug) are implemented as non-members, and deleted.  */
>  
> +  /* Convert this object to a std::string, using MAPPING as
> +     enumerator-to-string mapping array.  This is not meant to be
> +     called directly.  Instead, enum_flags specializations should have
> +     their own to_string function wrapping this one, thus hidding the
> +     mapping array from callers.  */
> +
> +  template<size_t N>
> +  std::string
> +  to_string (const string_mapping (&mapping)[N]) const
> +  {
> +    enum_type flags = raw ();
> +    std::string res = hex_string (flags);
> +    res += " [";
> +
> +    bool need_space = false;
> +    for (const auto &entry : mapping)
> +      {
> +	if ((flags & entry.flag) != 0)
> +	  {
> +	    /* Do op~ in the underlying type, because if enum_type's
> +	       underlying type is signed, op~ won't be defined for
> +	       it.  */
> +	    flags &= (enum_type) ~(underlying_type) entry.flag;

This line gives me (GCC 12):

  CXX    unittests/enum-flags-selftests.o
In file included from /home/simark/src/binutils-gdb/gdb/defs.h:65,
                 from /home/simark/src/binutils-gdb/gdb/unittests/enum-flags-selftests.c:20:
/home/simark/src/binutils-gdb/gdb/../gdbsupport/enum-flags.h: In instantiation of ‘std::string enum_flags<E>::to_string(const string_mapping (&)[N]) const [with long unsigned int N = 2; E = selftests::enum_flags_tests::test_uflag; std::string = std::__cxx11::basic_string<char>]’:
/home/simark/src/binutils-gdb/gdb/unittests/enum-flags-selftests.c:388:26:   required from here
/home/simark/src/binutils-gdb/gdb/../gdbsupport/enum-flags.h:219:19: error: invalid conversion from ‘unsigned int’ to ‘enum_flags<selftests::enum_flags_tests::test_uflag>::enum_type’ {aka ‘selftests::enum_flags_tests::test_uflag’} [-fpermissive]
  219 |             flags &= (enum_type) ~(underlying_type) entry.flag;
      |             ~~~~~~^~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
      |                   |
      |                   unsigned int

This builds though:

   flags = (enum_type) (flags & (enum_type) (~(underlying_type) entry.flag));

I'm not sure why, the operator & between an enum_type and an enum_type
is supposed to return an enum_type already.

Simon