I wonder if you can do some damage with scatter/gather ops within a VM, such that each fetch is a TLB miss inside the VM, and every table walk fetch is a TLB miss outside of the VM (which gets you up to 24 "fetches per fetch").
Chris usually takes an educational angle, I don't think this is LLM-generated content at all it's just his style. I highly encourage watching some of his DefCon or BlackHat talks, they're fun!
Bus cycles can be arbitrarily long on any processor that has memory cycles with a hand shake requiring an ack, with no timeout.
E.g. we can build a board around a MC68000 where we make it lock up forever in a bus cycle, waiting for a DTACK that doesn't arrive.
Some early microprocessors had clocked bus cycles without handshaking. They would put out an address on some address lines and signal some line together with a read/write indication, and then expect the transfer to be completed within some clock cycles. If nothing is attached to the address, they would read whatever values are on the bus, like maybe all 1's if it is an open drain system that requires the transmitting device to pull to ground to indicate zero.
I'd say that kind of thing belongs to a hall of shame; it requires software hacks to interface with anything that can't keep up with the prescribed bus cycle.
Ultimately... failure modes must arise because naive gate design simulation models can't determine issues in a computationally feasible time frame.
The consequences of "fixing" CDC prone design flaws makes a processor many times slower (8 to 16 times slower on my dumb attempt), and develops weird alien design features very different from Von Neumann architectures.
I thought I remembered reading somewhere re: the 8086 microcode disassembly that NOP, which is encoded as XCHG AX,AX actually does run the XCHG microcode and uses an internal scratchpad register to do the exchange.
As specified by the spec, it arguably increments RIP by one.
The actual typical hardware implementation just fetches the next 16-32 bytes from icache, shifts it to the correct alignment, and slams it into a bunch of parallel decoders which each attempts to decode one x86 instruction per byte.
The next cycle, the first 1-6 non-overlapping valid instructions are accepted into a queue for further decoding. The NOP almost certainly takes up space in this queue.
At no point does RIP get incremented by one. There isn't even a single physical RIP register to increment, the CPU is "executing" dozens or even hundreds of RIPs in parallel.
It's not an implementation detail, because the decoder runs before the execution of every instruction. If we're going to say that NOP increments IP by one, then we should also say that ADD "stores in dst the addition of src and dst, as well as incrementing IP by the length of the instruction", and JMP imm "increments JMP by imm + the length of the instruction".
I'm not disputing the total effect. I'm asking if you'd rather describe ADD and JMP in this manner, in order to say that NOP does not in fact do nothing.
This author also has other things like: A compiler that emits only `mov` instructions and another compiler that deliberately messes with the control flow so that, if disassembled, common debuggers will draw symbols like skulls or threats. https://github.com/xoreaxeaxeax/repsych
Also came up with the original ..cantor.dust.. binary visualization tool, which is a tool I never used directly but Chris' presentation of it in 2012 is still one of the coolest talks I've ever seen.
Very cool! Also, huh interesting. I’ve used rdtsc to measure cycle diffs but had no idea its execution takes that long. Is that common across architectures?
There are two aspects to compute performance: latency and throughput.
There has been a ton of work to improve throughput, but that's often come at the cost of latency, because a great technique to improve throughput is batching, to avoid the per-task overhead cost.
We've also added many layers of abstraction.
A seminal example is Dan Luu's "computer latency" table (2017) https://danluu.com/input-lag/, which measures the time between a keypress and visible changes on the screen, and wherein an Apple IIe has latency 5x lower than a Lenovo X1 Carbon on Windows.
From some perspective, you could say that the Lenovo on Windows example is an impressive feat of engineering, considering all the subsystems involved (USB, interrupt dispatcher, input layer, windowing system, double-buffering...)
This throughput-over-latency tradeoff can be seen across the entire computer landscape design.
The new mspaint fucked, then unfucked, then refucked my decades-old muscle memory of Win+R mspaint Enter Ctrl+E 1 Tab 1 Enter Ctrl+V to open Paint, resize canvas to minimum, then paste from clipboard. When you press Ctrl+E now, the Units control is selected by default, for some completely asinine reason!!!
It is literally impossible to respond to input in 10ms on most platforms, for various reasons. The USB input lag of 12-30ms and the 60Hz refresh rate of most monitors being just the first two.
The basic advice regarding response times has been about the same for thirty years [Miller 1968; Card et al. 1991]:
- 0.1 second is about the limit for having the user feel that the system is reacting instantaneously, meaning that no special feedback is necessary except to display the result.
- 1.0 second is about the limit for the user's flow of thought to stay uninterrupted, even though the user will notice the delay. Normally, no special feedback is necessary during delays of more than 0.1 but less than 1.0 second, but the user does lose the feeling of operating directly on the data.
- 10 seconds is about the limit for keeping the user's attention focused on the dialogue. For longer delays, users will want to perform other tasks while waiting for the computer to finish, so they should be given feedback indicating when the computer expects to be done. Feedback during the delay is especially important if the response time is likely to be highly variable, since users will then not know what to expect.
That’s the prevailing statistic, but your original number isn’t that wrong either:
Humans can perceive much smaller latencies.
If you look at the Card & Miller reference, at least some humans can perceive differences in ~50ms vs 100ms latencies when typing (in my limited testing, it’s likely you can!). There’s some newer research I don’t have handy that I believe found error rates decreased and NSAT improved until around at least 30ms (if not 20ms).
On that note, humans can definitely distinguish 60hz vs 120hz reliably (about 8ms faster per frame).
Even faster: with a reference (eg when dragging on a touchscreen), humans can distinguish down to at least 1ms vs 10ms of latency: https://m.youtube.com/watch?v=vOvQCPLkPt4
And you can probably distinguish metronomes that are off by about 1-2ms. Much smaller for other things (like metronomes that slightly slower or faster than one another).
60Hz monitors definitely prevent it, but I'm pretty sure USB lag is far less than 12-30ms. My USB mouse can make a round-trip to a remote server faster than that.
Sorry for nitpicking but the logical inverse of this statement is the stronger and more appropriate version:
If some app does not respond in 10ms or less, it is not interactive.
It'd be really interesting to see whether the winning (losing?) instructions/strategies would be different on other architectures. At least right now the top spot (`fxrstor64` on MMIO, starve PCIe) seems relatively architecture-independent, but maybe something about MMIO ordering rules on e.g. POWER would be different enough to change that -- or perhaps open up new avenues?
I wonder what the actual limit on this `fxrstor64` is right now. If you can stall the PCIe bus for that long, then why not indefinitely? Certainly there's no forward progress guarantee here.
There’s definitely strategies here; A lot of the floating point operations use subnormals, and a lot of the worst instructions are slowed down by really, really fucking with MMIO.
submitted by two different people, both with year+ old accounts and decent karma. i dont think either is the author. the other submitter probably read this one, looked at the github, saw something else cool and posted it. (i almost did the same, but bookmarked it instead)
I wonder if you can do some damage with scatter/gather ops within a VM, such that each fetch is a TLB miss inside the VM, and every table walk fetch is a TLB miss outside of the VM (which gets you up to 24 "fetches per fetch").
Related, and linked in the readme: https://github.com/xoreaxeaxeax/smiiiiiiiiiiiiiiii (using the slow instructions to break SMI)
I wish they would just explain it in normal terms instead of this nasty LLM "engaging blog post" style
Chris usually takes an educational angle, I don't think this is LLM-generated content at all it's just his style. I highly encourage watching some of his DefCon or BlackHat talks, they're fun!
I looked at both links and don't see anything weird or annoying, and I hate overblown styles myself.
Bus cycles can be arbitrarily long on any processor that has memory cycles with a hand shake requiring an ack, with no timeout.
E.g. we can build a board around a MC68000 where we make it lock up forever in a bus cycle, waiting for a DTACK that doesn't arrive.
Some early microprocessors had clocked bus cycles without handshaking. They would put out an address on some address lines and signal some line together with a read/write indication, and then expect the transfer to be completed within some clock cycles. If nothing is attached to the address, they would read whatever values are on the bus, like maybe all 1's if it is an open drain system that requires the transmitting device to pull to ground to indicate zero.
I'd say that kind of thing belongs to a hall of shame; it requires software hacks to interface with anything that can't keep up with the prescribed bus cycle.
In general, more complex processors have latency issues, and in some ways modern chips have actually become worse with each design iteration.
https://en.wikipedia.org/wiki/Metastability_(electronics)
Ultimately... failure modes must arise because naive gate design simulation models can't determine issues in a computationally feasible time frame.
The consequences of "fixing" CDC prone design flaws makes a processor many times slower (8 to 16 times slower on my dumb attempt), and develops weird alien design features very different from Von Neumann architectures.
This is why we can't have nice things. =3
It says in the rules
> Trapped/emulated/virtualized instructions may only time the trap, not the handler.
But I feel like that 12ms write to an ACPI IO port at current leaderboard position 8 is probably trapping to SMM and being handled there.
Nop should be #1, because it is infinitely slow for what it does. ;)
It increments rip by one.
I thought I remembered reading somewhere re: the 8086 microcode disassembly that NOP, which is encoded as XCHG AX,AX actually does run the XCHG microcode and uses an internal scratchpad register to do the exchange.
There were several NOPs - XCHG BX,BX and so on. Those were taken later to be prefixes for new classes of opcodes.
No, that's done by the decoder. It actually does nothing.
The decoder is an implementation detail that is a subcomponent of NOP; GP was right, and your correction isn't.
As specified by the spec, it arguably increments RIP by one.
The actual typical hardware implementation just fetches the next 16-32 bytes from icache, shifts it to the correct alignment, and slams it into a bunch of parallel decoders which each attempts to decode one x86 instruction per byte.
The next cycle, the first 1-6 non-overlapping valid instructions are accepted into a queue for further decoding. The NOP almost certainly takes up space in this queue.
At no point does RIP get incremented by one. There isn't even a single physical RIP register to increment, the CPU is "executing" dozens or even hundreds of RIPs in parallel.
It's not an implementation detail, because the decoder runs before the execution of every instruction. If we're going to say that NOP increments IP by one, then we should also say that ADD "stores in dst the addition of src and dst, as well as incrementing IP by the length of the instruction", and JMP imm "increments JMP by imm + the length of the instruction".
ADD does in fact do that.
I'm not disputing the total effect. I'm asking if you'd rather describe ADD and JMP in this manner, in order to say that NOP does not in fact do nothing.
It's still part of the instruction to increment it by one, as opposed to write a value or offset to it, as jump instructions do.
See my sibling response.
I mean....there are several architectures out there which has a nop that is a jump forward. Kind of a tree forest issue imho
I don't know about other architectures in as much detail. I know x86 NOP does nothing.
Strategy: nop does nothing. It opens the leaderboard accordingly.
Score: 1 cycles Time: 0 nanoseconds
It opens the leaderboard as #27, so in the last place.
It's a little faster than yep.
This author also has other things like: A compiler that emits only `mov` instructions and another compiler that deliberately messes with the control flow so that, if disassembled, common debuggers will draw symbols like skulls or threats. https://github.com/xoreaxeaxeax/repsych
He also bruteforced the entire opcode space to find undocumented instructions (sandsifter).
Also came up with the original ..cantor.dust.. binary visualization tool, which is a tool I never used directly but Chris' presentation of it in 2012 is still one of the coolest talks I've ever seen.
[0] https://github.com/Battelle/cantordust
[1] https://www.youtube.com/watch?v=4bM3Gut1hIk
Does that mean Chris Domas is ready for his next adventure?
what im seeing from this chart is that we should be using the nop instruction for everything
Well the best code is no code. Nop could be second best though.
Instructions unclear. Set the NX bit to ensure no code, and got a general protection fault.
Related, somehow: Core War [0].
[0]: https://en.wikipedia.org/wiki/Core_War
Very cool! Also, huh interesting. I’ve used rdtsc to measure cycle diffs but had no idea its execution takes that long. Is that common across architectures?
The cycle count for RDTSC is ~25 cycles on Skylake-era microarchitectures. The 49 number shown in the OP seems off.
AFAIK it acts as some kind of execution barrier, to give meaningful timing.
Isn't that rdtscp ( https://www.felixcloutier.com/x86/rdtscp )?
A great resource for any performance deoptimization.
It’s crazy how computers still seem to get perceivably slow every few years, given how many instructions can be executed in 1ms. Shameful, even..
What’s that law called about programmers wasting all the compute on abstraction?
There are two aspects to compute performance: latency and throughput.
There has been a ton of work to improve throughput, but that's often come at the cost of latency, because a great technique to improve throughput is batching, to avoid the per-task overhead cost.
We've also added many layers of abstraction.
A seminal example is Dan Luu's "computer latency" table (2017) https://danluu.com/input-lag/, which measures the time between a keypress and visible changes on the screen, and wherein an Apple IIe has latency 5x lower than a Lenovo X1 Carbon on Windows.
From some perspective, you could say that the Lenovo on Windows example is an impressive feat of engineering, considering all the subsystems involved (USB, interrupt dispatcher, input layer, windowing system, double-buffering...)
This throughput-over-latency tradeoff can be seen across the entire computer landscape design.
Idk man, my computers don't seem to get any slower over time -- no upgrades to any components, either.
Does responsiveness get better or worse with updates, in general?
Neither? I haven't noticed a perceivable change in a long time.
Wirth's Law I believe.
And remember to call him by name, not by value!
The new windows notepad is a disgrace
The new mspaint fucked, then unfucked, then refucked my decades-old muscle memory of Win+R mspaint Enter Ctrl+E 1 Tab 1 Enter Ctrl+V to open Paint, resize canvas to minimum, then paste from clipboard. When you press Ctrl+E now, the Units control is selected by default, for some completely asinine reason!!!
That's pretty cool. I've never needed to do that because Paint remembers the last canvas size set manually.
You aren't alone in being frustrated by decades of muscle memory being disrespected by MSFT.
i’m on LTSC for this (and many other) reasons lol! don’t touch my mspaint and notepad
Andy and Bill's Law
The OS should do less not more
vi should be a kernel-level system call, tunable with sysctls. For agentic management.
Huh? A millisecond is an eternity!
I remember reading once somewhere:
If some app responds in 10ms or less, it is INTERACTIVE.
makes you think.
It is literally impossible to respond to input in 10ms on most platforms, for various reasons. The USB input lag of 12-30ms and the 60Hz refresh rate of most monitors being just the first two.
I stand corrected.
I looked it up and it is .1 seconds (100ms)
The basic advice regarding response times has been about the same for thirty years [Miller 1968; Card et al. 1991]:
- 0.1 second is about the limit for having the user feel that the system is reacting instantaneously, meaning that no special feedback is necessary except to display the result.
- 1.0 second is about the limit for the user's flow of thought to stay uninterrupted, even though the user will notice the delay. Normally, no special feedback is necessary during delays of more than 0.1 but less than 1.0 second, but the user does lose the feeling of operating directly on the data.
- 10 seconds is about the limit for keeping the user's attention focused on the dialogue. For longer delays, users will want to perform other tasks while waiting for the computer to finish, so they should be given feedback indicating when the computer expects to be done. Feedback during the delay is especially important if the response time is likely to be highly variable, since users will then not know what to expect.
from Jakob Nielsen:
https://www.nngroup.com/articles/response-times-3-important-...
less readable but the original paper:
https://www.yusufarslan.net/sites/yusufarslan.net/files/uplo...
That’s the prevailing statistic, but your original number isn’t that wrong either:
Humans can perceive much smaller latencies.
If you look at the Card & Miller reference, at least some humans can perceive differences in ~50ms vs 100ms latencies when typing (in my limited testing, it’s likely you can!). There’s some newer research I don’t have handy that I believe found error rates decreased and NSAT improved until around at least 30ms (if not 20ms).
On that note, humans can definitely distinguish 60hz vs 120hz reliably (about 8ms faster per frame).
Even faster: with a reference (eg when dragging on a touchscreen), humans can distinguish down to at least 1ms vs 10ms of latency: https://m.youtube.com/watch?v=vOvQCPLkPt4
And you can probably distinguish metronomes that are off by about 1-2ms. Much smaller for other things (like metronomes that slightly slower or faster than one another).
This is a special interest of mine XD
60Hz monitors definitely prevent it, but I'm pretty sure USB lag is far less than 12-30ms. My USB mouse can make a round-trip to a remote server faster than that.
this is also why 60 hz refresh rate is all but dead outside of console gaming (not to mention 1000hz poll rate devices being the norm)
Sorry for nitpicking but the logical inverse of this statement is the stronger and more appropriate version: If some app does not respond in 10ms or less, it is not interactive.
Sahil
just curious, how much do these actually discover useful practices or pitfalls, on top of just being for fun?
Oh wow, glad to see Chris Domas active online again!
It'd be really interesting to see whether the winning (losing?) instructions/strategies would be different on other architectures. At least right now the top spot (`fxrstor64` on MMIO, starve PCIe) seems relatively architecture-independent, but maybe something about MMIO ordering rules on e.g. POWER would be different enough to change that -- or perhaps open up new avenues?
I wonder what the actual limit on this `fxrstor64` is right now. If you can stall the PCIe bus for that long, then why not indefinitely? Certainly there's no forward progress guarantee here.
Using MMIO is cheating and makes the results very boring.
It would be much more interesting to know the results if you're only allowed to use main memory.
There’s definitely strategies here; A lot of the floating point operations use subnormals, and a lot of the worst instructions are slowed down by really, really fucking with MMIO.
This is neat!
Seems like spam from this creator since there are two things on the front page?
submitted by two different people, both with year+ old accounts and decent karma. i dont think either is the author. the other submitter probably read this one, looked at the github, saw something else cool and posted it. (i almost did the same, but bookmarked it instead)