On 7 November 2016 at 23:36, John Steele Scott <toojays@xxxxxxxxxxx> wrote:
> On 04/11/16 22:32, John Steele Scott wrote:
>> Hi,
>>
>> A colleague of mine noticed that switching from -std=c++11 to -std=c++14
>> dramatically increased the size of our binary.
>>
>> After some investigation, this seems to be due to debuginfo.
>>
>> Consider the following simple example, compiled four different ways:
>>
>> jscott@citra:~/src/debuginfo-bloat$ cat simple.cpp
>> #include <string>
>>
>> size_t length (const std::string &data)
>> {
>> size_t len = 0;
>> for (auto iter = data.cbegin(); iter != data.cend(); iter++)
>> {
>> len++;
>> }
>>
>> return len;
>> }
>> jscott@citra:~/src/debuginfo-bloat$ g++-6 -gdwarf-4 -std=c++11 -c simple.cpp -o simple-11.o
>> jscott@citra:~/src/debuginfo-bloat$ g++-6 -gdwarf-4 -std=c++14 -c simple.cpp -o simple-14.o
>> jscott@citra:~/src/debuginfo-bloat$ g++-6 -gdwarf-4 -std=c++11 -c simple.cpp -o simple-11-g1.o -g1
>> jscott@citra:~/src/debuginfo-bloat$ g++-6 -gdwarf-4 -std=c++14 -c simple.cpp -o simple-14-g1.o -g1
>> jscott@citra:~/src/debuginfo-bloat$ ls -l simple*o
>> -rw-rw-r-- 1 jscott jscott 6904 Nov 4 22:25 simple-11-g1.o
>> -rw-rw-r-- 1 jscott jscott 60192 Nov 4 22:25 simple-11.o
>> -rw-rw-r-- 1 jscott jscott 6904 Nov 4 22:25 simple-14-g1.o
>> -rw-rw-r-- 1 jscott jscott 129376 Nov 4 22:25 simple-14.o
>>
>>
>> So with the default -g2, the C++14 output is almost double that of C++11. But
>> with -g1 they are the same.
>>
>> Can someone tell me why this is so? Am I getting much better debuginfo for those
>> bytes?
>
> I did some further digging with dwarfdump, and the increase in this case is due
> to std::literals::string_literals. If I ifdef those operators out (at the end of
> basic_string.h), the C++11 and C++14 sizes are the same.
I was going to suggest that this is almost certainly due to additional
declarations in the headers with C++14 is used, not a difference in
dwarf output caused by the -std options.
> I guess these operators cause GCC to think that the wchar_t, char16_t, char32_t
> versions of basic_string are used, and so it must emit debuginfo for them. But
> I'm not using the operators, therefore these classes are not used.
>
> Sounds like a GCC bug to me - it shouldn't emit debuginfo for types referenced
> by an inline function if that function is never used.
There might be some reason the debuginfo is useful in that case, but I
suggest opening a bug report so it can be assessed properly. Thanks.
