`vector_algorithms.cpp`: Remove the distinction between SSE2 and SSE4.2

[StephanTLavavej]

`vector_algorithms.cpp`: Remove the distinction between SSE2 and SSE4.2

[AlexGuteniev]

In this case some algorithms that avoid SSE4.2 may start using it. This allows simplifying/optimizing bitsett vectorization, also maybe some of reversing algos coud be direct pshufb

What is the fraction of SSE4.2 but not AVX?

[StephanTLavavej]

In this case some algorithms that avoid SSE4.2 may start using it.

Ah, even better! 😻

What is the fraction of SSE4.2 but not AVX?

In theory I could get accurate numbers, but I'd have to ask around through several people. The 2024-02 Steam Hardware Survey suggests a lower bound - it says that 0.46% of CPUs don't support SSE4.2 (which indicates that it's a vaguely reasonable lower bound for the numbers across all CPUs that we target, not just performance-minded gamers, given the similarity to the number I heard), while 6.75% don't support AVX2. I'd guess the actual number for us is in the range of 10-20%, so that AVX2 optimizations benefit the vast majority of users, but that we won't be able to assume its existence for a decade.

[AlexGuteniev]

Oh, I see there are still a lot of machines with AVX and not AVX2...
I wanted to know if the SSE code path is still useful at all, but looks like it is.
(It is possible to rewrite some of AVX2 algorithms to use AVX, but let's not do that for various reasons, including the reason of this issue)

[jovibor]

Oh, I see there are still a lot of machines with AVX and not AVX2...

Exactly.
Ideally vector_algorithms should provide something like (IMO):

if (Has_AVX2()) {
...
} else if (Has_AVX()) {
...
} else if (Has_SSE()) { //All SSEs (2, 3, 4.*).
...
} else { //Scalar.
...
}

AVX512 is out of this equation for the next decade I believe.

[StephanTLavavej]

Adding codepaths to distinguish AVX1 from AVX2 raises the same sort of hazards that I'm concerned about with SSE2 versus SSE4.2. Although the AVX1/AVX2 delta is maybe 4% of processors, I think the risk isn't worth it.

[jovibor]

Although the AVX1/AVX2 delta is maybe 4% of processors, I think the risk isn't worth it.

Then, am I right that your suggestion is:

//vector_algorithms.cpp

if (Has_AVX()) { //Exactly AVX1 and AVX2.
...
} else if (Has_SSE()) { //All SSEs (2, 3, 4.*).
...
} else { //Scalar.
...
}

[StephanTLavavej]

Yes, and we already have these functions (they are properly named _Use_avx2() and _Use_sse42()), so we just need to fuse the _Use_sse2() codepaths:

STL/stl/src/vector_algorithms.cpp

Lines 25 to 39 in be81252

	bool _Use_avx2() noexcept {
	return __isa_enabled & (1 << __ISA_AVAILABLE_AVX2);
	}
	bool _Use_sse42() noexcept {
	return __isa_enabled & (1 << __ISA_AVAILABLE_SSE42);
	}
	bool _Use_sse2() noexcept {
	#ifdef _M_IX86
	return __isa_enabled & (1 << __ISA_AVAILABLE_SSE2);
	#else
	return true;
	#endif
	}