runtime: swiss map SIMD implementation for non-amd64

[prattmic]

https://go.dev/cl/626277 (in 1.24) added SIMD implementations for swissmap (still 8-byte) control word operations for amd64. We could do similar for other architectures, if the SIMD implementations improve performance. ARM64 is the obvious target here given its wide use.

[Clement-Jean]

I'm interested in helping on this. I'll dig in the AMD64 implementation and try to come up with the ARM64 one

[Clement-Jean]

I poked around a little bit today, but I can seem to make the intrinsics work...

First of all, I'm surprised that I cannot simply set this alias to work on ARM64 (add sys.ArchARM64):

alias("internal/runtime/maps", "bitsetFirst", "internal/runtime/sys", "TrailingZeros64", sys.ArchAMD64)

This is weird to me because TrailingZeros64 is defined for ARM64 and this returns an error related to WASM when compiling:

panic: runtime error: index out of range [-16384]

goroutine 1 [running]:
cmd/internal/obj/wasm.init.0()
       /go/src/cmd/internal/obj/wasm/wasmobj.go:91 +0x1f4
panic: runtime error: index out of range [-16384]

goroutine 1 [running]:
cmd/internal/obj/wasm.init.0()
        /go/src/cmd/internal/obj/wasm/wasmobj.go:91 +0x1f4
panic: runtime error: index out of range [-16384]

goroutine 1 [running]:
cmd/internal/obj/wasm.init.0()
        /go/src/cmd/internal/obj/wasm/wasmobj.go:91 +0x1f4
panic: runtime error: index out of range [-16384]

goroutine 1 [running]:
cmd/internal/obj/wasm.init.0()
       /go/src/cmd/internal/obj/wasm/wasmobj.go:91 +0x1f4
go: error obtaining buildID for go tool compile: exit status 2
go: error obtaining buildID for go tool compile: exit status 2
go: error obtaining buildID for go tool compile: exit status 2
go: error obtaining buildID for go tool compile: exit status 2
panic: runtime error: index out of range [-16384]

goroutine 1 [running]:
cmd/internal/obj/wasm.init.0()
        /go/src/cmd/internal/obj/wasm/wasmobj.go:91 +0x1f4
go: error obtaining buildID for go tool compile: exit status 2
panic: runtime error: index out of range [-16384]

goroutine 1 [running]:
cmd/internal/obj/wasm.init.0()
       /go/src/cmd/internal/obj/wasm/wasmobj.go:91 +0x1f4
panic: runtime error: index out of range [-16384]

goroutine 1 [running]:
cmd/internal/obj/wasm.init.0()
        /go/src/cmd/internal/obj/wasm/wasmobj.go:91 +0x1f4
panic: runtime error: index out of range [-16384]

goroutine 1 [running]:
cmd/internal/obj/wasm.init.0()
        /go/src/cmd/internal/obj/wasm/wasmobj.go:91 +0x1f4
go: error obtaining buildID for go tool compile: exit status 2
go: error obtaining buildID for go tool compile: exit status 2
go: error obtaining buildID for go tool compile: exit status 2
go: internal error: Builder WorkDir unexpectedly changed from /var/folders/nv/rqg_nqy55pq_ky2j_34whq6c0000gn/T/go-build1659387489 to
panic: runtime error: index out of range [-16384]

goroutine 1 [running]:
cmd/internal/obj/wasm.init.0()
        go/src/cmd/internal/obj/wasm/wasmobj.go:91 +0x1f4
go tool dist: FAILED: go/pkg/tool/darwin_arm64/go_bootstrap install -a cmd/asm cmd/cgo cmd/compile cmd/link cmd/preprofile: exit status 1

maybe I forgot to regenerate something?

After that, I tried to make the bitsetLowestSet intrinsic work but ssa.OpEq64 doesn't seem to accept a constInt64. Could you guide me to where are the rules for instructions (like what they accept as params and at which position)? I remember there is such a table but I can't seem to find it...

[thepudds]

Hi @Clement-Jean, I think it is great that you are exploring this.

I wanted to point you to two optional resources that I think could be helpful if you don’t already know about them:

• The #contributors channel in Gopher Slack is a good place for both general and specific questions about contributing to Go.
• The #assembly channel in Gopher slack has friendly/knowledgeable people.

There is an invite app for access if you are interested.

There are also the mailing lists and other forums, and people will comment here as well, but I wanted to mention those two channels as options for informal or ad-hoc conversations that tend to be fairly helpful.

[Clement-Jean]

It seems that without the implementation of intrinsics, vector types, and register allocation for vectors, the ARM64 implementation is largely inefficient.

I tried both representation of the bitset (packed and unpacked) and my implementation is almost always 2 times slower than the current one on the BenchmarkMapAccessHit benchmark.

If anyone wants to check my implementation you can find it here (on branch simd_swissmap_arm64): https://github.com/Clement-Jean/go