Wasmtime GitHub notifications bot (Jul 20 2022 at 11:38): Wasmtime GitHub notifications bot (Jul 20 2022 at 11:38):afonso360 opened PR #4474 from x86-vex to main:
:wave: Hey,
Following the discussion in #4462, this PR adds a VEX encoder and lowering for SIMD
fmawith FMA extension instructions.The VEX Encoder has an API similar to the EVEX encoder, however instead of writing the instruction as the caller is calling the fields, we store them all and later write the full instruction in
encode.This is because the VEX instruction format is much less predictable than EVEX and I couldn't find a way to cleanly write the bit fields, since at any time we can switch from 2 to 3 byte prefixes.
@abrown mentioned in #4462, there are potentially some improvements that we may want to do.
With that in place, we would probably need think through how to emit AVX instructions; e.g., something like Avx512Opcode but perhaps we want to be able to decide on the VEX/EVEX encoding at a later time (?)
Right now this implements
AvxOpcodeand always lowers it in VEX encoding. However I like the idea of choosing the encoding based on what is available and most efficient. Suggestions here would be appreciated!Fixes #4462
Wasmtime GitHub notifications bot (Jul 20 2022 at 11:38):afonso360 edited PR #4474 from x86-vex to main:
:wave: Hey,
Following the discussion in #4462, this PR adds a VEX encoder and lowering for SIMD
fmawith FMA extension instructions.The VEX Encoder has an API similar to the EVEX encoder, however instead of writing the instruction as the caller is calling the fields, we store them all and later write the full instruction in
encode.This is because the VEX instruction format is much less predictable than EVEX and I couldn't find a way to cleanly write the bit fields, since at any time we can switch from 2 to 3 byte prefixes.
@abrown mentioned in #4462, there are potentially some improvements that we may want to do.
With that in place, we would probably need think through how to emit AVX instructions; e.g., something like Avx512Opcode but perhaps we want to be able to decide on the VEX/EVEX encoding at a later time (?)
Right now this implements
AvxOpcodeand always lowers it in VEX encoding. However I like the idea of choosing the encoding based on what is available and most efficient. Suggestions on how we would implement this would be appreciated!Fixes #4462
Wasmtime GitHub notifications bot (Jul 20 2022 at 11:48):afonso360 updated PR #4474 from x86-vex to main.
Wasmtime GitHub notifications bot (Jul 20 2022 at 12:28):afonso360 updated PR #4474 from x86-vex to main.
Wasmtime GitHub notifications bot (Jul 20 2022 at 12:29):afonso360 updated PR #4474 from x86-vex to main.
Wasmtime GitHub notifications bot (Jul 25 2022 at 18:25):abrown created PR review comment:
How come we need these moves?
Wasmtime GitHub notifications bot (Jul 25 2022 at 18:25):abrown submitted PR review.
Wasmtime GitHub notifications bot (Jul 25 2022 at 18:25):abrown submitted PR review.
Wasmtime GitHub notifications bot (Jul 25 2022 at 18:29):cfallin submitted PR review.
Wasmtime GitHub notifications bot (Jul 25 2022 at 18:29):cfallin created PR review comment:
Ah, I suspect this may be due to using a "mod" operand rather than a separate src and dest with dest tied to src via a reuse-constraint. @afonso360 would you mind matching how the other instructions work in that way? Then you can assert that src1 == dst during emission and the emission code remains unchanged, but it makes the usage here simpler.
Wasmtime GitHub notifications bot (Jul 25 2022 at 18:30):afonso360 created PR review comment:
They aren't actually emitted, but
xisn't writable, so we need a reg that is, and sincevfmadd213pdmodifies the first reg we also need the value ofxto be in that register. I'm not sure if this is the best way to achieve that.
Wasmtime GitHub notifications bot (Jul 25 2022 at 18:30):afonso360 submitted PR review.
Wasmtime GitHub notifications bot (Jul 25 2022 at 18:32):afonso360 edited PR review comment.
Wasmtime GitHub notifications bot (Jul 25 2022 at 19:16):afonso360 updated PR #4474 from x86-vex to main.
Wasmtime GitHub notifications bot (Jul 25 2022 at 19:25):afonso360 updated PR #4474 from x86-vex to main.
Wasmtime GitHub notifications bot (Jul 25 2022 at 20:42):cfallin submitted PR review.
Wasmtime GitHub notifications bot (Jul 25 2022 at 20:42):cfallin created PR review comment:
The order of these uses of
allocsneeds to match the order in which registers are given to theOperandCollectoringet_operandsbelow. (I know, it's a terrible footgun... we hope to make this a bit more ergonomic later!)
Wasmtime GitHub notifications bot (Jul 25 2022 at 20:42):cfallin submitted PR review.
Wasmtime GitHub notifications bot (Jul 25 2022 at 20:55):afonso360 updated PR #4474 from x86-vex to main.
Wasmtime GitHub notifications bot (Jul 25 2022 at 20:56):afonso360 updated PR #4474 from x86-vex to main.
Wasmtime GitHub notifications bot (Jul 25 2022 at 20:57):afonso360 created PR review comment:
Thanks! I'm still not too comfortable with this.
Wasmtime GitHub notifications bot (Jul 25 2022 at 20:57):afonso360 submitted PR review.
Wasmtime GitHub notifications bot (Jul 25 2022 at 21:27):cfallin submitted PR review.
Wasmtime GitHub notifications bot (Jul 25 2022 at 21:30):cfallin has enabled auto merge for PR #4474.
Wasmtime GitHub notifications bot (Jul 25 2022 at 22:01):cfallin merged PR #4474.
Last updated: Nov 22 2024 at 16:03 UTC