Merge pull request #523 from koute/main_runtime_simd

Runtime backend autodetection

.github/workflows/rust.yml

@@ -55,19 +55,19 @@ jobs:
       - run: cargo build --target thumbv7em-none-eabi --release
       - run: cargo build --target thumbv7em-none-eabi --release --features serde
 
-  build-simd-nightly:
-    name: Build simd backend (nightly)
+  test-simd-native:
+    name: Test simd backend (native)
     runs-on: ubuntu-latest
     steps:
     - uses: actions/checkout@v3
     - uses: dtolnay/rust-toolchain@nightly
-    # Build with AVX2 features, then with AVX512 features
     - env:
-        RUSTFLAGS: '--cfg curve25519_dalek_backend="simd" -C target_feature=+avx2'
-      run: cargo build --target x86_64-unknown-linux-gnu
-    - env:
-        RUSTFLAGS: '--cfg curve25519_dalek_backend="simd" -C target_feature=+avx512ifma'
-      run: cargo build --target x86_64-unknown-linux-gnu
+        # This will:
+        #   1) build all of the x86_64 SIMD code,
+        #   2) run all of the SIMD-specific tests that the test runner supports,
+        #   3) run all of the normal tests using the best available SIMD backend.
+        RUSTFLAGS: '-C target_cpu=native'
+      run: cargo test --features simd --target x86_64-unknown-linux-gnu
 
   test-simd-avx2:
     name: Test simd backend (avx2)
@@ -76,8 +76,10 @@ jobs:
     - uses: actions/checkout@v3
     - uses: dtolnay/rust-toolchain@stable
     - env:
-        RUSTFLAGS: '--cfg curve25519_dalek_backend="simd" -C target_feature=+avx2'
-      run: cargo test --target x86_64-unknown-linux-gnu
+        # This will run AVX2-specific tests and run all of the normal tests
+        # with the AVX2 backend, even if the runner supports AVX512.
+        RUSTFLAGS: '-C target_feature=+avx2'
+      run: cargo test --no-default-features --features alloc,precomputed-tables,zeroize,simd_avx2 --target x86_64-unknown-linux-gnu
 
   build-docs:
     name: Build docs
@@ -131,12 +133,7 @@ jobs:
     - uses: dtolnay/rust-toolchain@nightly
       with:
         components: clippy
-    - env:
-        RUSTFLAGS: '--cfg curve25519_dalek_backend="simd" -C target_feature=+avx2'
-      run: cargo clippy --target x86_64-unknown-linux-gnu
-    - env:
-        RUSTFLAGS: '--cfg curve25519_dalek_backend="simd" -C target_feature=+avx512ifma'
-      run: cargo clippy --target x86_64-unknown-linux-gnu
+    - run: cargo clippy --target x86_64-unknown-linux-gnu
 
   rustfmt:
     name: Check formatting
@@ -162,9 +159,7 @@ jobs:
     - uses: dtolnay/[email protected]
     - run: cargo build --no-default-features --features serde
     # Also make sure the AVX2 build works
-    - env:
-        RUSTFLAGS: '--cfg curve25519_dalek_backend="simd" -C target_feature=+avx2'
-      run: cargo build --target x86_64-unknown-linux-gnu
+    - run: cargo build --target x86_64-unknown-linux-gnu
 
   bench:
     name: Check that benchmarks compile

Cargo.toml

@@ -27,7 +27,6 @@ rustdoc-args = [
     "--html-in-header", "docs/assets/rustdoc-include-katex-header.html",
     "--cfg", "docsrs",
 ]
-rustc-args = ["--cfg", "curve25519_dalek_backend=\"simd\""]
 features = ["serde", "rand_core", "digest", "legacy_compatibility"]
 
 [dev-dependencies]
@@ -54,15 +53,29 @@ digest = { version = "0.10", default-features = false, optional = true }
 subtle = { version = "2.3.0", default-features = false }
 serde = { version = "1.0", default-features = false, optional = true, features = ["derive"] }
 zeroize = { version = "1", default-features = false, optional = true }
+unsafe_target_feature = { version = "= 0.1.1", optional = true }
+
+[target.'cfg(target_arch = "x86_64")'.dependencies]
+cpufeatures = "0.2.6"
 
 [target.'cfg(curve25519_dalek_backend = "fiat")'.dependencies]
 fiat-crypto = "0.1.19"
 
 [features]
-default = ["alloc", "precomputed-tables", "zeroize"]
+default = ["alloc", "precomputed-tables", "zeroize", "simd"]
 alloc = ["zeroize?/alloc"]
 precomputed-tables = []
 legacy_compatibility = []
 
+# Whether to allow the use of the AVX2 SIMD backend.
+simd_avx2 = ["unsafe_target_feature"]
+
+# Whether to allow the use of the AVX512 SIMD backend.
+# (Note: This requires Rust nightly; on Rust stable this feature will be ignored.)
+simd_avx512 = ["unsafe_target_feature"]
+
+# A meta-feature to allow all SIMD backends to be used.
+simd = ["simd_avx2", "simd_avx512"]
+
 [profile.dev]
 opt-level = 2

Makefile

@@ -1,6 +1,5 @@
 FEATURES := serde rand_core digest legacy_compatibility
 
-export RUSTFLAGS := --cfg=curve25519_dalek_backend="simd"
 export RUSTDOCFLAGS := \
 	--cfg docsrs \
 	--html-in-header docs/assets/rustdoc-include-katex-header.html

README.md

@@ -53,6 +53,9 @@ curve25519-dalek = "4.0.0-rc.2"
 | `alloc`            |    ✓     | Enables Edwards and Ristretto multiscalar multiplication, batch scalar inversion, and batch Ristretto double-and-compress. Also enables `zeroize`. |
 | `zeroize`          |    ✓     | Enables [`Zeroize`][zeroize-trait] for all scalar and curve point types. |
 | `precomputed-tables` |    ✓     | Includes precomputed basepoint multiplication tables. This speeds up `EdwardsPoint::mul_base` and `RistrettoPoint::mul_base` by ~4x, at the cost of ~30KB added to the code size. |
+| `simd_avx2`        |    ✓     | Allows the AVX2 SIMD backend to be used, if available. |
+| `simd_avx512`      |    ✓     | Allows the AVX512 SIMD backend to be used, if available. |
+| `simd`             |    ✓     | Allows every SIMD backend to be used, if available. |
 | `rand_core`        |          | Enables `Scalar::random` and `RistrettoPoint::random`. This is an optional dependency whose version is not subject to SemVer. See [below](#public-api-semver-exemptions) for more details. |
 | `digest`           |          | Enables `RistrettoPoint::{from_hash, hash_from_bytes}` and `Scalar::{from_hash, hash_from_bytes}`. This is an optional dependency whose version is not subject to SemVer. See [below](#public-api-semver-exemptions) for more details. |
 | `serde`            |          | Enables `serde` serialization/deserialization for all the point and scalar types. |
@@ -95,30 +98,26 @@ See tracking issue: [curve25519-dalek/issues/521](https://github.com/dalek-crypt
 
 Curve arithmetic is implemented and used by selecting one of the following backends:
 
-| Backend            | Implementation                                             | Target backends             |
-| :---               | :---                                                       | :---                        |
-| `[default]`        | Serial formulas                                            | `u32` <br/> `u64`           |
-| `simd`             | [Parallel][parallel_doc], using Advanced Vector Extensions | `avx2` <br/> `avx512ifma`   |
-| `fiat`             | Formally verified field arithmetic from [fiat-crypto]      | `fiat_u32` <br/> `fiat_u64` |
+| Backend            | Implementation                                                | Target backends             |
+| :---               | :---                                                          | :---                        |
+| `[default]`        | Automatic runtime backend selection (either serial or SIMD)   | `u32` <br/> `u64` <br/> `avx2` <br/> `avx512` |
+| `fiat`             | Formally verified field arithmetic from [fiat-crypto]         | `fiat_u32` <br/> `fiat_u64` |
 
-To choose a backend other than the `[default]` serial backend, set the
+To choose a backend other than the `[default]` backend, set the
 environment variable:
 ```sh
 RUSTFLAGS='--cfg curve25519_dalek_backend="BACKEND"'
 ```
-where `BACKEND` is `simd` or `fiat`. Equivalently, you can write to
+where `BACKEND` is `fiat`. Equivalently, you can write to
 `~/.cargo/config`:
 ```toml
 [build]
 rustflags = ['--cfg=curve25519_dalek_backend="BACKEND"']
 ```
 More info [here](https://doc.rust-lang.org/cargo/reference/config.html#buildrustflags).
 
-The `simd` backend requires extra configuration. See [the SIMD
-section](#simd-target-backends).
-
 Note for contributors: The target backends are not entirely independent of each
-other. The `simd` backend directly depends on parts of the the `u64` backend to
+other. The SIMD backend directly depends on parts of the the `u64` backend to
 function.
 
 ## Word size for serial backends
@@ -137,7 +136,7 @@ RUSTFLAGS='--cfg curve25519_dalek_bits="SIZE"'
 where `SIZE` is `32` or `64`. As in the above section, this can also be placed
 in `~/.cargo/config`.
 
-**NOTE:** The `simd` backend CANNOT be used with word size 32.
+**NOTE:** Using a word size of 32 will automatically disable SIMD support.
 
 ### Cross-compilation
 
@@ -152,18 +151,19 @@ $ cargo build --target i686-unknown-linux-gnu
 
 ## SIMD target backends
 
-Target backend selection within `simd` must be done manually by setting the
-`RUSTFLAGS` environment variable to one of the below options:
+The SIMD target backend selection is done automatically at runtime depending
+on the available CPU features, provided the appropriate feature flag is enabled.
 
-| CPU feature | `RUSTFLAGS`                     | Requires nightly? |
-| :---        | :---                            | :---              |
-| avx2        | `-C target_feature=+avx2`       | no                |
-| avx512ifma  | `-C target_feature=+avx512ifma` | yes               |
+You can also specify an appropriate `-C target_feature` to build a binary
+which assumes the required SIMD instructions are always available.
 
-Or you can use `-C target_cpu=native` if you don't know what to set.
+| Backend | Feature flag  | `RUSTFLAGS`                               | Requires nightly? |
+| :---    | :---          | :---                                      | :---              |
+| avx2    | `simd_avx2`   | `-C target_feature=+avx2`                 | no                |
+| avx512  | `simd_avx512` | `-C target_feature=+avx512ifma,+avx512vl` | yes               |
 
-The AVX512 backend requires Rust nightly. If enabled and when compiled on a non-nightly
-compiler it will fall back to using the AVX2 backend.
+The AVX512 backend requires Rust nightly. When compiled on a non-nightly
+compiler it will always be disabled.
 
 # Documentation
 
@@ -243,16 +243,16 @@ The implementation is memory-safe, and contains no significant
 `unsafe` code.  The SIMD backend uses `unsafe` internally to call SIMD
 intrinsics.  These are marked `unsafe` only because invoking them on an
 inappropriate CPU would cause `SIGILL`, but the entire backend is only
-compiled with appropriate `target_feature`s, so this cannot occur.
+invoked when the appropriate CPU features are detected at runtime, or
+when the whole program is compiled with the appropriate `target_feature`s.
 
 # Performance
 
 Benchmarks are run using [`criterion.rs`][criterion]:
 
 ```sh
 cargo bench --features "rand_core"
-# Uses avx2 or ifma only if compiled for an appropriate target.
-export RUSTFLAGS='--cfg curve25519_dalek_backend="simd" -C target_cpu=native'
+export RUSTFLAGS='-C target_cpu=native'
 cargo +nightly bench --features "rand_core"
 ```
 
@@ -294,7 +294,7 @@ universe's beauty, but also his deep hatred of the Daleks. Rusty destroys the
 other Daleks and departs the ship, determined to track down and bring an end
 to the Dalek race.*
 
-`curve25519-dalek` is authored by Isis Agora Lovecruft and Henry de Valence. 
+`curve25519-dalek` is authored by Isis Agora Lovecruft and Henry de Valence.
 
 Portions of this library were originally a port of [Adam Langley's
 Golang ed25519 library](https://github.com/agl/ed25519), which was in

build.rs

@@ -27,6 +27,13 @@ fn main() {
     {
         println!("cargo:rustc-cfg=nightly");
     }
+
+    let rustc_version = rustc_version::version().expect("failed to detect rustc version");
+    if rustc_version.major == 1 && rustc_version.minor <= 64 {
+        // Old versions of Rust complain when you have an `unsafe fn` and you use `unsafe {}` inside,
+        // so for those we want to apply the `#[allow(unused_unsafe)]` attribute to get rid of that warning.
+        println!("cargo:rustc-cfg=allow_unused_unsafe");
+    }
 }
 
 // Deterministic cfg(curve25519_dalek_bits) when this is not explicitly set.