Gate AVX-512 behind chacha20_avx512 cfg

Author: caelunshun

chacha20/Cargo.toml

@@ -3,7 +3,7 @@ name = "chacha20"
 version = "0.10.0-rc.2"
 authors = ["RustCrypto Developers"]
 edition = "2024"
-rust-version = "1.89"
+rust-version = "1.85"
 documentation = "https://docs.rs/chacha20"
 readme = "README.md"
 repository = "https://github.com/RustCrypto/stream-ciphers"
@@ -49,6 +49,7 @@ rustdoc-args = ["--cfg", "docsrs"]
 [lints.rust.unexpected_cfgs]
 level = "warn"
 check-cfg = [
+    'cfg(chacha20_avx512)',
     'cfg(chacha20_force_soft)',
     'cfg(chacha20_force_sse2)',
     'cfg(chacha20_force_avx2)',

chacha20/README.md

@@ -32,6 +32,7 @@ work on stable Rust with the following `RUSTFLAGS`:
 - `x86` / `x86_64`
   - `avx2`: (~1.4cpb) `-Ctarget-cpu=haswell -Ctarget-feature=+avx2`
   - `sse2`: (~1.6cpb) `-Ctarget-feature=+sse2` (on by default on x86 CPUs)
+  - `avx512`: `-Ctarget-feature=+avx512f,+avx512vl --cfg chacha20_avx512` requires Rust 1.89+
 - `aarch64`
   - `neon` (~2-3x faster than `soft`) requires Rust 1.61+ and the `neon` feature enabled
 - Portable

chacha20/src/backends.rs

@@ -7,14 +7,15 @@ cfg_if! {
         pub(crate) mod soft;
     } else if #[cfg(any(target_arch = "x86", target_arch = "x86_64"))] {
         cfg_if! {
-            if #[cfg(chacha20_force_avx512)] {
+            if #[cfg(all(chacha20_avx512, chacha20_force_avx512))] {
                 pub(crate) mod avx512;
             } else if #[cfg(chacha20_force_avx2)] {
                 pub(crate) mod avx2;
             } else if #[cfg(chacha20_force_sse2)] {
                 pub(crate) mod sse2;
             } else {
                 pub(crate) mod soft;
+                #[cfg(chacha20_avx512)]
                 pub(crate) mod avx512;
                 pub(crate) mod avx2;
                 pub(crate) mod sse2;

chacha20/src/lib.rs

@@ -185,9 +185,9 @@ cfg_if! {
         type Tokens = ();
     } else if #[cfg(any(target_arch = "x86", target_arch = "x86_64"))] {
         cfg_if! {
-            if #[cfg(chacha20_force_avx512)] {
+            if #[cfg(all(chacha20_avx512, chacha20_force_avx512))] {
                 #[cfg(not(all(target_feature = "avx512f", target_feature = "avx512vl")))]
-                compile_error!("You must enable `avx512f` target feature with \
+                compile_error!("You must enable `avx512f` and `avx512vl` target features with \
                     `chacha20_force_avx512` configuration option");
                 type Tokens = ();
             } else if #[cfg(chacha20_force_avx2)] {
@@ -201,10 +201,14 @@ cfg_if! {
                     `chacha20_force_sse2` configuration option");
                 type Tokens = ();
             } else {
+                #[cfg(chacha20_avx512)]
                 cpufeatures::new!(avx512_cpuid, "avx512f", "avx512vl");
                 cpufeatures::new!(avx2_cpuid, "avx2");
                 cpufeatures::new!(sse2_cpuid, "sse2");
+                #[cfg(chacha20_avx512)]
                 type Tokens = (avx512_cpuid::InitToken, avx2_cpuid::InitToken, sse2_cpuid::InitToken);
+                #[cfg(not(chacha20_avx512))]
+                type Tokens = (avx2_cpuid::InitToken, sse2_cpuid::InitToken);
             }
         }
     } else {
@@ -259,8 +263,10 @@ impl<R: Rounds, V: Variant> ChaChaCore<R, V> {
                         let tokens = ();
                     } else if #[cfg(chacha20_force_sse2)] {
                         let tokens = ();
-                    } else {
+                    } else if #[cfg(chacha20_avx512)] {
                         let tokens = (avx512_cpuid::init(), avx2_cpuid::init(), sse2_cpuid::init());
+                    } else {
+                        let tokens = (avx2_cpuid::init(), sse2_cpuid::init());
                     }
                 }
             } else {
@@ -306,7 +312,7 @@ impl<R: Rounds, V: Variant> StreamCipherCore for ChaChaCore<R, V> {
                 f.call(&mut backends::soft::Backend(self));
             } else if #[cfg(any(target_arch = "x86", target_arch = "x86_64"))] {
                 cfg_if! {
-                    if #[cfg(chacha20_force_avx512)] {
+                    if #[cfg(all(chacha20_avx512, chacha20_force_avx512))] {
                         unsafe {
                             backends::avx512::inner::<R, _, V>(&mut self.state, f);
                         }
@@ -319,12 +325,19 @@ impl<R: Rounds, V: Variant> StreamCipherCore for ChaChaCore<R, V> {
                             backends::sse2::inner::<R, _, V>(&mut self.state, f);
                         }
                     } else {
+                        #[cfg(chacha20_avx512)]
                         let (avx512_token, avx2_token, sse2_token) = self.tokens;
+                        #[cfg(not(chacha20_avx512))]
+                        let (avx2_token, sse2_token) = self.tokens;
+
+                        #[cfg(chacha20_avx512)]
                         if avx512_token.get() {
                             unsafe {
                                 backends::avx512::inner::<R, _, V>(&mut self.state, f);
                             }
-                        } else if avx2_token.get() {
+                            return;
+                        }
+                        if avx2_token.get() {
                             unsafe {
                                 backends::avx2::inner::<R, _, V>(&mut self.state, f);
                             }