Hacker News Reader

Comment by thesnide | Hacker News Reader

I'm wondering if we'll see some riscv extensions specifically designed to improve x86/amd64 emulation, such as what the M1 and other did.

View on HN

Some upcoming chips are supposed to support switching individual processes to x86's "TSO" memory model. That might be the most significant extension that Apple has for x86 emulation: it allows eliding all memory fence instructions used to adapt to the weaker memory model.

LoongArch could have instructions that emulate specific x86 behaviour and flags, but there is practically no documentation available.

Is it just "some upcoming chips" inventing their own extensions? Or is this a standardized ARM extension?

Basically, will writing against these upcoming chips mean writing one implementation for Qualcomm, one implementation for Rockchip, one implementation for Samsung, etc? Or will it just require one implementation for the standard ARM "switch to total store ordering memory model" extension

>Is it just "some upcoming chips" inventing their own extensions? Or is this a standardized ARM extension?

RISC-V has an official ratified extension for TSO, and a work-in-progress one for dynamic switching between RISC-V's standard memory model and TSO.

I'm sorry, I meant RISC-V, not ARM. So far the RISC-V standard has specified behaviour under the TSO memory model and a flag in the ELF header for code that has been compiled for TSO. There is not yet any ratified extension for dynamic switching of memory model but I'd expect anything vendor-specific to be wrapped behind a Linux syscall.

Reply on news.ycombinator.com