3 ms·

Calling it “X86S” was setting it up to fail for one. We need a new ISA, not a reinvention of the last.

by cedws 2y ago

jauntywundrkind 2y ago
Why? Casting Doubt (friend of Fear and Uncertainty) is easy & viral. It's so unclear that it matters at all though. We have incredibly awesome front ends that reprocess the x86 instructions into more tangible CPU-ish work, micro-opd and macro-opd and op caching. Yes, it is certainly painful to do. But the front end doesn't seem to be super energy intensive and isn't that big. Ultimately if we did try to build a core where the instruction set was directly executable we would be fixed in place, wouldn't be able to adapt our execution units well. In some ways, having the execution units decoupled from the instruction set allows flexibility & adaption, allows independent evolution, that may have been a strong boon to us. Just saying we should build a better ISA sounds a bit pompous, an immature/negative intolerance to nuamce or demand of perfection - a tear it down mentality - when the complexities of life & machine architecture demand iterative adaptation & change.
- cedws 2y ago
  If you're a manufacturer that's going to put the industry through the pain of a new ISA, why not start on a clean slate and wipe out 40 years of preconceptions and mistakes? Why not dare to reimagine? If you rework the foundations you can unlock huge leaps that were previously unimaginable. Apple could have kept using Intel chips and still had MacBooks flying off the shelves for years to come, but they took a risk by tearing it all up and their product is better than ever. (yes, I am aware Apple did not invent their own ISA.)
  - jauntywundrkind 2y ago
    > that's going to put the industry through the pain of a new ISA, X86s is supposed to radically simplify the pain, get rid of a very arcane ratcheting up of antiquated unused modes. It's supposed to decrease pain. And simplify implementation too. We don't dare to reimagine because it's not clear what ISA would be better or what the wins would really be. Folks like Apple turn to existing ancient instruction sets like ARM rather than reinvent because it works and the ISA doesn't really seem to have that much influence on the core design & performance.
    - hakfoo 2y ago
      The risk with X86S was that people didn't really know what the corner cases were. I'd assume if you start with a fairly vanilla system, all modern x86-64 code, then yeah it's going to work fine. But I could imagine some of the modes and instructions that might get yeeted could cause unforeseen consequences on more "delicate" software. Obviously this means stuff like "386-only binary applications from 1991" but I'd wonder if some of the weird modes are intentionally triggered by things like copy-protection software to make them less easily reverse-engineered.

Calling it “X86S” was setting it up to fail for one. We need a new ISA, not a reinvention of the last.

by cedws 2y ago

Calling it “X86S” was setting it up to fail for one. We need a new ISA, not a reinvention of the last.

Why? Casting Doubt (friend of Fear and Uncertainty) is easy & viral. It's so unclear that it matters at all though. We have incredibly awesome front ends that reprocess the x86 instructions into more tangible CPU-ish work, micro-opd and macro-opd and op caching. Yes, it is certainly painful to do. But the front end doesn't seem to be super energy intensive and isn't that big. Ultimately if we did try to build a core where the instruction set was directly executable we would be fixed in place, wouldn't be able to adapt our execution units well. In some ways, having the execution units decoupled from the instruction set allows flexibility & adaption, allows independent evolution, that may have been a strong boon to us. Just saying we should build a better ISA sounds a bit pompous, an immature/negative intolerance to nuamce or demand of perfection - a tear it down mentality - when the complexities of life & machine architecture demand iterative adaptation & change.

cedws 2y ago
If you're a manufacturer that's going to put the industry through the pain of a new ISA, why not start on a clean slate and wipe out 40 years of preconceptions and mistakes? Why not dare to reimagine? If you rework the foundations you can unlock huge leaps that were previously unimaginable. Apple could have kept using Intel chips and still had MacBooks flying off the shelves for years to come, but they took a risk by tearing it all up and their product is better than ever. (yes, I am aware Apple did not invent their own ISA.)
- jauntywundrkind 2y ago
  > that's going to put the industry through the pain of a new ISA, X86s is supposed to radically simplify the pain, get rid of a very arcane ratcheting up of antiquated unused modes. It's supposed to decrease pain. And simplify implementation too. We don't dare to reimagine because it's not clear what ISA would be better or what the wins would really be. Folks like Apple turn to existing ancient instruction sets like ARM rather than reinvent because it works and the ISA doesn't really seem to have that much influence on the core design & performance.
  - hakfoo 2y ago
    The risk with X86S was that people didn't really know what the corner cases were. I'd assume if you start with a fairly vanilla system, all modern x86-64 code, then yeah it's going to work fine. But I could imagine some of the modes and instructions that might get yeeted could cause unforeseen consequences on more "delicate" software. Obviously this means stuff like "386-only binary applications from 1991" but I'd wonder if some of the weird modes are intentionally triggered by things like copy-protection software to make them less easily reverse-engineered.