So Intel probably shoved popcnt into the same category to keep the processor design simple
In the processor design I work on, we do register dependency checks by partitioning all instructions into a set of "timing classes" and checking the dispatch delay needed between dependent register producers and consumers across all possible timing class pairs. The delays vary depending on available forwarding networks, resource conflicts, etc. Often times we groups instructions into sub optimal timing classes to simplify other parts of the design or just to make the dispatch logic simpler.
Intel's x86 core is waaaaay more complicated than the core I work on and has far more instructions, so I it's probably safe to say that they make these suboptimal classifications often. I strongly suspect that the false dependency was intentional and not a "hardware bug" as some of the StackOverflow comments seem to suggest.
I wouldn't classify it as intentional nor a "bug"; probably it's more of an oversight, as it's mentioned in the article that AMD's CPUs don't have this issue. Intel should definitely be made aware of this.
We can only speculate, but it's likely that Intel has the same handling for a lot two-operand instructions. Common instructions like add, sub take two operands both of which are inputs. So Intel probably shoved popcnt into the same category to keep the processor design simple.
On the other hand, MOV doesn't read both operands either.
'icpc' (the Intel C++ compiler) has equal performance for both of the test cases, and it did choose to use different registers for each call. But it's not clear if that's by design or by chance. In some ways, that's the boring part. The interesting part (to me) is that both tests are much faster than either version with g++.
Here's icpc 14.0.3 vs g++ 4.8.1 on a Sandy Bridge E5-1620 @ 3.60GHz and a Haswell i7-4770 CPU @ 3.40GHz.
That would be incredible if true! But I think it's a bug, since the inner loop looks far too short and doesn't seem to be repeating the popcnt's. I'm not sure yet if it's a problem with the compiler or if the test case is abusing something undefined.
OK, it looks like 'icpc' has decided that it would be fastest to invert the two loops: popcnt() once, then repeat the addition 10000 times. I'm neither a language lawyer nor a friend of C++, so I'll refrain to trying to decide whether this is a legal optimization. But a liberal sprinkling of 'volatile' makes it do what was obviously intended. After this, the speeds are more comparable, although 'icpc' retains a small (but much more plausible) lead:
The other test I did was checking what Intel's IACA (a wonderful optimization tool that you really should be using if you are not already) thought about the g++ loop. It did _not_ notice the false dependency, and said the loops should take the same amount of time. Do this suggest that the Intel compiler is just getting lucky, or that Intel doesn't have great internal communication between teams?
It's definitely legal, the buffer is created with operator new and is not made visible to any other code, so the compiler knows that it can't change within the loop.
Even if the problem isn't there, it's really easy to fix in that layer: just insert an instruction before popcnt that kills the value in the destination register, and there won't be anything to wait for. Intel does regular microcode updates to fix this sort of thing, so I would anticipate seeing this one get fixed in the not-too-distant future.
TLDR: Headline (and indeed bulk of article) is phantom symptom. True cause is register allocator behavior.
Specifically, allocator's handling of an instruction with a false dependency on register that's written to, coupled with multiple compilers being unaware of the false dependency.
Maybe one should add, that (as much I understood) it is a problem of the processor handling one specific (and rare) instruction. It does assume register dependencies that do not exist. It was shown, that AMD does not have this behavior. And it shows, that today's processors are enormous complex beasts.
The problem with the compilers was, that they where not aware of this behavior and thus generated sub-optimal code for this situation ... but compiler builders are also mere humans.
This is why micro-benchmarking is Russian roulette.
When you distill a loop until you're finding the exact bottleneck in the system (pipelining, branch prediction, etc) you need to be very very careful you're measuring what you think you are. Otherwise you'll end up in this situation where you're benchmarking a compiler...
I suppose similarly related to this, when I was keeping track of synchronization between two cooperative simulation threads running at different frequencies, I had a 64-bit signed integer: chip A would add chip_B_frequency * chip_A_cycles_executed; and chip B would subtract chip_A_frequency * chip_B_cycles_executed. If the value was >=0, chip A was ahead and would switch to B; and if the value was <0, chip B was ahead and would switch to A.
I ended up getting a noticeable speed boost just by using sync += (uint32_t)clocks * (uint64_t)frequency; ... just a simple 32-bit x 64-bit multiply was quite a bit faster than a 64-bit x 64-bit multiply. (One had to be 64-bit to prevent the multiplication from overflowing, as one value was in the MHz range and the other could be up to ~2000 or so.)
I've observed this on both AMD and Intel amd64 CPUs. Not sure how that'd hold up on other CPUs. As always though, profile your code first, and only consider these types of tricks in hot code areas.
It should be noted that using 64-bit operands, even in 64-bit mode, incurs an extra penalty of 1 byte per instruction, for the REX prefix. The same applies to using the extended registers (the uncreatively-named "r8" through "r15".) This is very much not noticeable for microbenchmarks, where all the code of a loop fits in the cache, but for bigger ones, the effects of icache misses can become quite significant. A smaller instruction sequence that is slower than a larger one when microbenchmarked can become much faster once that code is benchmarked as part of a whole system.
Intel's x86 core is waaaaay more complicated than the core I work on and has far more instructions, so I it's probably safe to say that they make these suboptimal classifications often. I strongly suspect that the false dependency was intentional and not a "hardware bug" as some of the StackOverflow comments seem to suggest.