df98494a30
I filed #49331 about five years ago, suggesting that the compiler should re-associate some computations to support ILP. There was general consensus that it was a decent idea, and I believe that Ryan Berger may have attempted it, but so far, nothing has landed. For md5 in particular, which was the motivating example, we can get the desired parallelism by adjusting the source code. So do that. We can always revert if/when the compiler pass lands. Each MD5 step is of the form arg0 = arg1 + RotateLeft32(F(arg1,arg2,arg3) + arg0 + x + k, s) where arg1 is the output of the previous step. The generated code has the form a+(b+(c+d)) rather than (a+b)+(c+d), which has a long serial dependency chain. This is a significant source of the performance difference between the generic Go code and the handwritten assembly. Adjust the generator to impose the association we desire. Group the arg1-independent terms as F + (arg0+x+k). This lets the compiler schedule the inner sum in parallel with F, trimming the chain to a single add after F is calculated. Add two further expression tweaks. In round 2, use the direct selection form (x&z)|(y&^z) instead of the XOR-trick form ((x^y)&z)^y. The direct form puts arg1 on a 2-op path to F instead of 3, which reduces serialization on the critical path and shortens variable lifetimes. On arm64, both forms are 3 logical ops. On default amd64, this adds an instruction, but it's off the critical path. And BMI1 is available at GOAMD64 >= v3. In round 3, associate arg1^arg2^arg3 as arg1^(arg2^arg3), again for ILP. Before this commit, asm was about ~30% faster than purego on my arm64 machine, and ~37% on the amd64 machine I have access to (no GOAMD64). This commit narrows the gap to ~13% and ~20%, respectively. I have follow-up work planned that shrinks the gap further and should enable deletion of the md5 assembly. Updates #49248 Updates #49331 Change-Id: Ia317e20cd73c88c4fb9303b31f5b6b441bfa6b63 Reviewed-on: https://go-review.googlesource.com/c/go/+/801240 LUCI-TryBot-Result: golang-scoped@luci-project-accounts.iam.gserviceaccount.com <golang-scoped@luci-project-accounts.iam.gserviceaccount.com> Reviewed-by: Russ Cox <rsc@golang.org> Reviewed-by: Filippo Valsorda <filippo@golang.org> Reviewed-by: Michael Pratt <mpratt@google.com>
The Go Programming Language
Go is an open source programming language that makes it easy to build simple, reliable, and efficient software.
Gopher image by Renee French, licensed under Creative Commons 4.0 Attribution license.
Our canonical Git repository is located at https://go.googlesource.com/go. There is a mirror of the repository at https://github.com/golang/go.
Unless otherwise noted, the Go source files are distributed under the BSD-style license found in the LICENSE file.
Download and Install
Binary Distributions
Official binary distributions are available at https://go.dev/dl/.
After downloading a binary release, visit https://go.dev/doc/install for installation instructions.
Install From Source
If a binary distribution is not available for your combination of operating system and architecture, visit https://go.dev/doc/install/source for source installation instructions.
Contributing
Go is the work of thousands of contributors. We appreciate your help!
To contribute, please read the contribution guidelines at https://go.dev/doc/contribute.
Note that the Go project uses the issue tracker for bug reports and proposals only. See https://go.dev/wiki/Questions for a list of places to ask questions about the Go language.