search

ptitSeb / box64

Box64 - Linux Userspace x86_64 Emulator with a twist, targeted at ARM64 Linux devices
https://box86.org
MIT License
3.73k stars 267 forks source link

[RV64_DYNAREC] Fixed vector infra #1705

Closed ksco closed 2 months ago

ksco commented 2 months ago

For vlen>128, we cannot use load/store whole vector instructions! Sorry for this mistake. This PR replaced all the whole vector load/stores with the normal ones, which requires the vtype to be set correctly using SET_ELEMENT_WIDTH().

ksco commented 2 months ago

Related to this PR:

I think there are certain limitations in the current vector infrastructure -- mainly for opcodes that are essentially element-width-agnostic.

For example, MOVDQA can be implemented with vle8.v, vle16.v, vle32.v, and vle64.v, it will be good if we know the predecessor's element width in compile time to choose the best one.

Previously I implemented MOVDQA with "load whole vector", but that unfortunately only works on vlen=128.

ptitSeb commented 2 months ago

Related to this PR:

I think there are certain limitations in the current vector infrastructure -- mainly for opcodes that are essentially element-width-agnostic.

For example, MOVDQA can be implemented with vle8.v, vle16.v, vle32.v, and vle64.v, it will be good if we know the predecessor's element width in compile time to choose the best one.

Previously I implemented MOVDQA with "load whole vector", but that unfortunately only works on vlen=128.

You can use a ecanism similar to float/i64/double of Float, but SSE handling is quite different, so an other mecanism should be used.