Skip to content

Possible performance loss with f32 arithmetic #91447

Closed
@Yuri6037

Description

@Yuri6037

I've tried, out of curiosity, a floating point arithmetic test and found quite a big difference between C++ and Rust.

The code used in rust

pub struct Stats
{
    x: f32,
    y: f32,
    z: f32
}

pub fn sum(a: &Stats, b: &Stats) -> Stats
{
    Stats {
        x: a.x + b.x,
        y: a.y + b.y,
        z: a.z + b.z
    }
}

The code used in C++

struct Stats
{
    float x;
    float y;
    float z;
};

Stats sum(const Stats &a, const Stats &b)
{
    return Stats {
        a.x + b.x,
        a.y + b.y,
        a.z + b.z
    };
}

Here is a link to a godbolt for side-by-side comparision of assembly output: https://godbolt.org/z/dqc4b74rv

Rust seem to absolutely want the floats back into e* registers instead of keeping them in xmm registers, C++ leaves them into the xmm registers. In some cases it might more advantageous to leave the floats in xmm registers for future operations on them rather then passing them back into the e* registers.

Metadata

Metadata

Assignees

No one assigned

    Labels

    A-LLVMArea: Code generation parts specific to LLVM. Both correctness bugs and optimization-related issues.A-codegenArea: Code generationA-floating-pointArea: Floating point numbers and arithmeticC-bugCategory: This is a bug.I-slowIssue: Problems and improvements with respect to performance of generated code.T-compilerRelevant to the compiler team, which will review and decide on the PR/issue.

    Type

    No type

    Projects

    No projects

    Milestone

    No milestone

    Relationships

    None yet

    Development

    No branches or pull requests

    Issue actions