<iostream>: Bad performance of wcout operator<< compared to wprintf

[bernd5]

If I compile the following code:

for(auto idx = 0; idx < 10; idx++){
        const wchar_t* a = L"this is a test\n";      
        std::wcout << a;
}

it needs round about 18 milliseconds during runtime.

But if I use C functions like:

for(auto idx = 0; idx < 10; idx++){
        const wchar_t* a = L"this is a test\n";      
        wprintf(L"%s", a);
}

It needs only 5 milliseconds!
Why is std::cout / wcout so slow? I would assume that the C function should be slower because it has to analyze the format string, but...

I tried to optimize it by calling first: std::ios::sync_with_stdio(false);
But this seems to be completely ignored (I found no reference to _Sync...)?

[bernd5]

link to an old article:
Link

[bernd5]

with this call it becomes fast:
setvbuf(stdout, 0, _IOLBF, 4096);

Why is the default so bad?

[StephanTLavavej]

Why is the default so bad?

As mentioned in my decade-old reply to that Connect bug you found, this is a CRT limitation:

(As documented at https://docs.microsoft.com/en-us/cpp/c-runtime-library/reference/setvbuf?view=msvc-160 [updated link] , our C Standard Library implementation isn't capable of line buffering, only full buffering or no buffering. Paragraph 7.19.3/7 of the 1999 C Standard requires that "As initially opened, the standard error stream is not fully buffered; the standard input and standard output streams are fully buffered if and only if the stream can be determined not to refer to an interactive device." Since we don't have line buffering, we must default to no buffering.)

We changed the iostreams code in that decade - now, xsputn attempts to avoid character-by-character processing, but only for 1-byte characters, so your wchar_t repro doesn't benefit from that optimization:

STL/stl/inc/fstream

Lines 634 to 636 in 280347a

	} else { // non-chars always get element-by-element processing
	return _Mysb::xsputn(_Ptr, _Count);
	}

I'll leave this issue active for further investigation; it may be possible to extend this optimization to wchar_t.