cranelift / Issue #1350 Determine best way to encode SIMD... · git-cranelift

Stream: git-cranelift

Topic: cranelift / Issue #1350 Determine best way to encode SIMD...

GitHub (Jan 14 2020 at 23:50):

What is the feature or code improvement you would like to do in Cranelift?

I would like to implement the SIMD load_extend instructions.

What is the value of adding this in Cranelift?

This is necessary for Wasm SIMD spec compliance.

Do you have an implementation plan, and/or ideas for data structures or
algorithms to use?

I see that currently shared/instructions.rs includes instructions such as uload32 and sload32; one option is to add the additional instructions [u/s]load8x8, [u/s]load16x4, and [u/s]load32x4. Alternately, @bnjbvr and I had discussed using the IR->IR infrastructure of legalization to implement peephole optimizations; if this was in place and at the right level, I could translate the Wasm load8x8_s, e.g., to Cranelift's load.i8x8 and uextend and then write a peephole optimization to generate the appropriate x86 PMOVSXBW to encode both of these instructions (perhaps with a new Cranelift x86_pmovsxb instruction). Which approach is better.

GitHub (Jan 14 2020 at 23:51):

abrown commented on Issue #1350:

cc: @bnjbvr, @sunfishcode

GitHub (Jan 14 2020 at 23:51):

abrown edited Issue #1350:

What is the feature or code improvement you would like to do in Cranelift?

I would like to implement the SIMD load_extend instructions.

What is the value of adding this in Cranelift?

This is necessary for Wasm SIMD spec compliance.

Do you have an implementation plan, and/or ideas for data structures or algorithms to use?

I see that currently shared/instructions.rs includes instructions such as uload32 and sload32; one option is to add the additional instructions [u/s]load8x8, [u/s]load16x4, and [u/s]load32x4. Alternately, @bnjbvr and I had discussed using the IR->IR infrastructure of legalization to implement peephole optimizations; if this was in place and at the right level, I could translate the Wasm load8x8_s, e.g., to Cranelift's load.i8x8 and uextend and then write a peephole optimization to generate the appropriate x86 PMOVSXBW to encode both of these instructions (perhaps with a new Cranelift x86_pmovsxb instruction). Which approach is better.

GitHub (Feb 06 2020 at 19:38):

sunfishcode commented on Issue #1350:

Significant peephole-style optimization will be best done once https://github.com/bytecodealliance/cranelift/issues/1344 is in place and we can build up from there. So my suggestion here is to just add the additional instructions, [u/s]load8x8, and so on, for now, and then we can revisit them later once we're ready to do more general peephole optimizations.

GitHub (Feb 28 2020 at 23:28):

alexcrichton transferred Issue #1350:

What is the feature or code improvement you would like to do in Cranelift?

I would like to implement the SIMD load_extend instructions.

What is the value of adding this in Cranelift?

This is necessary for Wasm SIMD spec compliance.

Do you have an implementation plan, and/or ideas for data structures or algorithms to use?

I see that currently shared/instructions.rs includes instructions such as uload32 and sload32; one option is to add the additional instructions [u/s]load8x8, [u/s]load16x4, and [u/s]load32x4. Alternately, @bnjbvr and I had discussed using the IR->IR infrastructure of legalization to implement peephole optimizations; if this was in place and at the right level, I could translate the Wasm load8x8_s, e.g., to Cranelift's load.i8x8 and uextend and then write a peephole optimization to generate the appropriate x86 PMOVSXBW to encode both of these instructions (perhaps with a new Cranelift x86_pmovsxb instruction). Which approach is better.

Last updated: Nov 22 2024 at 17:03 UTC