ExifKey::Impl::tagName() is Slow

[eyesoftime]

Describe the bug

As the title says: std::string ExifKey::Impl::tagName() const is slow.

To Reproduce

I've been trying to track down why readMetadata is slow for my Sony a99ii ARW files. Slow as in order of magnitude to two slower than with JPEGs or TIFFs. Found one of the issues to be here:

exiv2/src/tags.cpp

Lines 182 to 184 in b917b34

	std::ostringstream os;
	os << "0x" << std::setw(4) << std::setfill('0') << std::right << std::hex << tag_;
	return os.str();

It gets executed about two thousand times accumulating over 100ms. Converting a short unsigned int to a four-digit hex string shouldn't be that big of an issue. Testing around I found that the following cuts the time about three orders of magnitude. Feel free to substitute those three lines with something like the following. There's likely an even more efficient way but at this point it hardly matters for now. I'm down with metadata loading from 180ms to about 19ms but I'll have to dig around some more to see if there are other fairly obvious issues.

Expected behavior

I propose something like this instead of the three lines linked above. I'll defer to your judgement on the Endianness issues if any.

constexpr static char m[] { "0123456789abcdef" };
char s[] { "0x0000" };
s[2] = m[(tag_ >> 12) & 0xf];
s[3] = m[(tag_ >> 8) & 0xf];
s[4] = m[(tag_ >> 4) & 0xf];
s[5] = m[tag_ & 0xf];
return s;

Desktop (please complete the following information):

• OS and version: Win10
• Exiv2 version and source: 0.28.3
• Compiler and version: Clang 19.1.7
• Compilation mode and/or compiler flags: debugoptimized meson build, using as a static library

[neheb]

How about

return stringFormat("0x{:04x}", tag_);

?

[kmilos]

How about

Should be good for main branch. We don't have neither libfmt nor std::format on 0.28.x, so we probably won't change anything there unfortunately....

[eyesoftime]

return stringFormat("0x{:04x}", tag_);

Sure, why not. Quick test shows it to be about 5x slower than what I wrote above but it does cut down two orders of magnitude so it should be much better already.