cmd/compile: missing inlining and/or SROA for SIMD code

[balasanjay]

What version of Go are you using (go version)?

$ go version
go version go1.26.3 linux/amd64

Does this issue reproduce with the latest release?

What operating system and processor architecture are you using (go env)?

go env Output

$ go env
AR='ar'
CC='cc'
CGO_CFLAGS='-O2 -g'
CGO_CPPFLAGS=''
CGO_CXXFLAGS='-O2 -g'
CGO_ENABLED='1'
CGO_FFLAGS='-O2 -g'
CGO_LDFLAGS='-O2 -g'
CXX='c++'
GCCGO='gccgo'
GO111MODULE=''
GOAMD64='v1'
GOARCH='amd64'
GOAUTH='netrc'
GOBIN=''
GOCACHE='/var/home/sjy/.cache/go-build'
GOCACHEPROG=''
GODEBUG=''
GOENV='/var/home/sjy/.config/go/env'
GOEXE=''
GOEXPERIMENT=''
GOFIPS140='off'
GOFLAGS=''
GOGCCFLAGS='-fPIC -m64 -pthread -Wl,--no-gc-sections -fmessage-length=0 -ffile-prefix-map=/tmp/go-build4263466051=/tmp/go-build -gno-record-gcc-switches'
GOHOSTARCH='amd64'
GOHOSTOS='linux'
GOINSECURE=''
GOMOD='/var/home/sjy/src/me/aegis/go.mod'
GOMODCACHE='/var/home/sjy/go/pkg/mod'
GONOPROXY=''
GONOSUMDB=''
GOOS='linux'
GOPATH='/var/home/sjy/go'
GOPRIVATE=''
GOPROXY='https://proxy.golang.org,direct'
GOROOT='/var/home/linuxbrew/.linuxbrew/Cellar/go/1.26.3/libexec'
GOSUMDB='sum.golang.org'
GOTELEMETRY='local'
GOTELEMETRYDIR='/var/home/sjy/.config/go/telemetry'
GOTMPDIR=''
GOTOOLCHAIN='auto'
GOTOOLDIR='/var/home/linuxbrew/.linuxbrew/Cellar/go/1.26.3/libexec/pkg/tool/linux_amd64'
GOVCS=''
GOVERSION='go1.26.3'
GOWORK=''
PKG_CONFIG='pkg-config'
uname -sr: Linux 6.19.12-200.fc43.x86_64
LSB Version:	n/a
Distributor ID:	Bluefin
Description:	Bluefin (Version: 43.20260505)
Release:	43
Codename:	Deinonychus
/lib64/libc.so.6: GNU C Library (GNU libc) stable release version 2.42.

What did you do?

I wrote some simd/archsimd code to implement the AEGISx2 encryption algorithm.

I got pretty good performance (encryption of a 65k blob with a throughput of 1.5GB/s), but I knew that the C and Zig implementations had achieved much higher performance.

I then manually inlined 2 functions (UpdateState128x2, and Enc128x2), and manually hoisted my working set of SIMD variables out of a struct into local variables (aka a manual SROA optimization): diff.

With that change, I got a significant boost in performance (encryption of a 65k blob with a throughput of 28.5GB/s).

What did you expect to see?

I expected the compiler to automatically inline these functions, apply SROA-based optimization and have the high performance with the straightforward code.

What did you see instead?

Required manual optimization.

---

Here are the instructions to reproduce:

go mod init example.com/repro

# Performance without manual inlining.
go get github.com/balasanjay/aegis@14a5c57a492ba15937c79798495a515f8ec87aea
GOAMD64=v3 GOEXPERIMENT=simd go test github.com/balasanjay/aegis/... -bench . -count=1

# Performance with manual inlining
go get github.com/balasanjay/aegis@3de416051d0e9a17a643a38cd7d686a7a5b199f3
GOAMD64=v3 GOEXPERIMENT=simd go test github.com/balasanjay/aegis/... -bench . -count=1

[balasanjay]

cc @JunyangShao

[gabyhelp]

Related Issues

• simd/archsimd: would intrinsics eliminate Go/assembly boundary overhead for SIMD prefilter? #77647 (closed)
• inlining reduces performance compared to manual inlining #35687 (closed)
• simd/archsimd: archsimd.LoadUint8x32Slice() emits nops #76968 (closed)
• cmd/objdump: doesn't properly disassemble code generated with simd/archsimd intrinsics #78221 (closed)
• cmd/compile: riscv performance degredation #74606 (closed)
• cmd/compile: doesn't inline basic funcs and doesn't optimize code after inlining #70235
• cmd/compile: inline is slower #51028 (closed)
• cmd/compile: inliner does not inline `binary.*Endian.Uint*` after many inlines in a big function #68081

_{(Emoji vote if this was helpful or unhelpful; more detailed feedback welcome in this discussion.)}

[JunyangShao]

Thanks for opening the issue!
I have a pending CL on the main branch: go.dev/cl/742701. Once it's reviewed and +2-ed I will submit it asap.