Find out what features are supported by what CPU

djc · February 8, 2021, 9:42am

I'm wondering about making better use of available CPU features I'm working on. While I currently use target-cpu=native in my development environment and in my staging environment (different kinds of macOS machines), I deploy into GCP on N2D instances (2nd-gen AMD Zen).

As I understand it, there's no good way to build docker images that abstract over this difference, so I'm wondering what the best way forward is. I found Using RUSTFLAGS - Rust SIMD Performance Guide which was quite helpful; it contains the text "The compiler will translate this into a list of target features.".

Is there some way of inspecting what features the compiler will enable for a given target CPU, so that I could define a list of "least common denominator" features available on both, say, znver2 and skylake? And does anyone have a sense of how impactful optimizations the compiler can make beyond the availability of features (for example, it looks like LLVM has a cost model for different processors)?

bjorn3 · February 8, 2021, 9:58am

Rustc directly gets this information from LLVM:

github.com

rust-lang/rust/blob/4940dd483a8448c0f1ef28d304fad88a9d983c4e/compiler/rustc_codegen_llvm/src/llvm_util.rs#L167


    supported_target_features(sess)
        .iter()
        .filter_map(
            |&(feature, gate)| {
                if sess.is_nightly_build() || gate.is_none() { Some(feature) } else { None }
            },
        )
        .filter(|feature| {
            let llvm_feature = to_llvm_feature(sess, feature);
            let cstr = CString::new(llvm_feature).unwrap();
            unsafe { llvm::LLVMRustHasFeature(target_machine, cstr.as_ptr()) }
        })
        .map(|feature| Symbol::intern(feature))
        .collect()
}
pub fn print_version() {
    let (major, minor, patch) = get_version();
    println!("LLVM version: {}.{}.{}", major, minor, patch);
}

For example for Zen2 LLVM has:

github.com

llvm/llvm-project/blob/c2c83e97c3ac98ddf5bd685cbfba3f620f59fa51/llvm/lib/Target/X86/X86.td#L1028-L1073


// AMD Zen Processors common ISAs
list<SubtargetFeature> ZNFeatures = [FeatureADX,
                                     FeatureAES,
                                     FeatureAVX2,
                                     FeatureBMI,
                                     FeatureBMI2,
                                     FeatureCLFLUSHOPT,
                                     FeatureCLZERO,
                                     FeatureCMOV,
                                     Feature64Bit,
                                     FeatureCMPXCHG16B,
                                     FeatureF16C,
                                     FeatureFMA,
                                     FeatureFSGSBase,
                                     FeatureFXSR,
                                     FeatureNOPL,
                                     FeatureLAHFSAHF,
                                     FeatureLZCNT,
                                     FeatureMMX,
                                     FeatureMOVBE,

This file has been truncated. show original

Skylake has the following in addition to what is available on Broadwell:

github.com

llvm/llvm-project/blob/c2c83e97c3ac98ddf5bd685cbfba3f620f59fa51/llvm/lib/Target/X86/X86.td#L652-L670


// Skylake
list<SubtargetFeature> SKLAdditionalFeatures = [FeatureAES,
                                                FeatureXSAVEC,
                                                FeatureXSAVES,
                                                FeatureCLFLUSHOPT,
                                                FeatureSGX];
list<SubtargetFeature> SKLTuning = [FeatureHasFastGather,
                                    FeatureMacroFusion,
                                    FeatureSlow3OpsLEA,
                                    FeatureSlowDivide64,
                                    FeatureFastScalarFSQRT,
                                    FeatureFastVectorFSQRT,
                                    FeatureFastSHLDRotate,
                                    FeatureFast15ByteNOP,
                                    FeatureFastVariableShuffle,
                                    FeaturePOPCNTFalseDeps,
                                    FeatureInsertVZEROUPPER];
list<SubtargetFeature> SKLFeatures =
  !listconcat(BDWFeatures, SKLAdditionalFeatures);

bluss · February 8, 2021, 10:18am

Be sure to look at the output of these two commands, it can be quite helpful, the target features are among the target cfgs being shown here:

rustc --print cfg
rustc --print cfg -C target-cpu=native

And the output can be different depending on stable vs nightly as well - with more cfgs being activated on nightly.

djc · February 8, 2021, 10:27am

Thanks, this is also very useful! Is this stuff documented anywhere in the rust-lang docs? And I presume CPU features do not vary across the OS part of a target triple, is that a safe assumption?

After comparing a list based on LLVM sources compared to what rustc --print cfg -C target-cpu=native gives me on skylake, quite a few things are missing (like x87, lahfsah or ermsb) from the output of the rustc command. I also see some inconsistency between the list printed by rustc for print --target-features vs print --cfg (like bmi vs bmi1 and pclmulqdq), any thoughts on that?

bluss · February 8, 2021, 1:03pm

What kind of inconsistency?

djc · February 8, 2021, 1:05pm

This seems weird: when passing the bmi1, pclmulqdq or rdrand options as a -Ctarget-feature value, I get a warning like '+bmi1' is not a recognized feature for this target (ignoring feature). Opened an issue about this: rustc --print cfg prints features that are invalid for `-Ctarget-feature` · Issue #81881 · rust-lang/rust · GitHub.

EdmundsEcho · February 8, 2021, 2:26pm

When I was going through the process I found it useful to ground/validate what the compiler was reporting with what I knew about my local CPU running OSX. In my case, I used

sysctl -a | grep machdep.cpu.features

... to get most of the features supported by my CPU. The full feature set I was expecting could all be found under the machdep.cpu register/setting.