Add Icelake avx512 features (#838)

* Add Icelake avx512 features

As documented in https://software.intel.com/sites/default/files/managed/c5/15//architecture-instruction-set-extensions-programming-reference.pdf

* Sort the avx512 feature checks by bit

* Unbreak macos

Force nightly.

Author: lu-zero

.github/workflows/main.yml

@@ -173,8 +173,12 @@ jobs:
       run: |
         curl https://sh.rustup.rs | sh -s -- -y --default-toolchain nightly
         echo "##[add-path]$HOME/.cargo/bin"
+        rustup update nightly --no-self-update
+        rustup default nightly
       if: matrix.os == 'macos-latest'
-    - run: rustup target add ${{ matrix.target }}
+    - run: |
+        rustup default nightly
+        rustup target add ${{ matrix.target }}
       if: "!endsWith(matrix.target, 'emulated')"
     - run: cargo generate-lockfile
 

crates/core_arch/tests/cpu-detection.rs

@@ -30,6 +30,23 @@ fn x86_all() {
         "avx512_vpopcntdq {:?}",
         is_x86_feature_detected!("avx512vpopcntdq")
     );
+    println!("avx512vbmi2 {:?}", is_x86_feature_detected!("avx512vbmi2"));
+    println!("avx512gfni {:?}", is_x86_feature_detected!("avx512gfni"));
+    println!("avx512vaes {:?}", is_x86_feature_detected!("avx512vaes"));
+    println!(
+        "avx512vpclmulqdq {:?}",
+        is_x86_feature_detected!("avx512vpclmulqdq")
+    );
+    println!("avx512vnni {:?}", is_x86_feature_detected!("avx512vnni"));
+    println!(
+        "avx512bitalg {:?}",
+        is_x86_feature_detected!("avx512bitalg")
+    );
+    println!("avx512bf16 {:?}", is_x86_feature_detected!("avx512bf16"));
+    println!(
+        "avx512vp2intersect {:?}",
+        is_x86_feature_detected!("avx512vp2intersect")
+    );
     println!("f16c: {:?}", is_x86_feature_detected!("f16c"));
     println!("fma: {:?}", is_x86_feature_detected!("fma"));
     println!("abm: {:?}", is_x86_feature_detected!("abm"));

crates/std_detect/src/detect/arch/x86.rs

@@ -135,6 +135,22 @@ features! {
     @FEATURE: #[unstable(feature = "stdsimd", issue = "27731")] avx512vpopcntdq: "avx512vpopcntdq";
     /// AVX-512 VPOPCNTDQ (Vector Population Count Doubleword and
     /// Quadword)
+    @FEATURE: #[unstable(feature = "stdsimd", issue = "27731")] avx512vbmi2: "avx512vbmi2";
+    /// AVX-512 VBMI2 (Additional byte, word, dword and qword capabilities)
+    @FEATURE: #[unstable(feature = "stdsimd", issue = "27731")] avx512gfni: "avx512gfni";
+    /// AVX-512 GFNI (Galois Field New Instruction)
+    @FEATURE: #[unstable(feature = "stdsimd", issue = "27731")] avx512vaes: "avx512vaes";
+    /// AVX-512 VAES (Vector AES instruction)
+    @FEATURE: #[unstable(feature = "stdsimd", issue = "27731")] avx512vpclmulqdq: "avx512vpclmulqdq";
+    /// AVX-512 VPCLMULQDQ (Vector PCLMULQDQ instructions)
+    @FEATURE: #[unstable(feature = "stdsimd", issue = "27731")] avx512vnni: "avx512vnni";
+    /// AVX-512 VNNI (Vector Neural Network Instructions)
+    @FEATURE: #[unstable(feature = "stdsimd", issue = "27731")] avx512bitalg: "avx512bitalg";
+    /// AVX-512 BITALG (Support for VPOPCNT[B,W] and VPSHUFBITQMB)
+    @FEATURE: #[unstable(feature = "stdsimd", issue = "27731")] avx512bf16: "avx512bf16";
+    /// AVX-512 BF16 (BFLOAT16 instructions)
+    @FEATURE: #[unstable(feature = "stdsimd", issue = "27731")] avx512vp2intersect: "avx512vp2intersect";
+    /// AVX-512 P2INTERSECT
     @FEATURE: #[unstable(feature = "stdsimd", issue = "27731")] f16c: "f16c";
     /// F16C (Conversions between IEEE-754 `binary16` and `binary32` formats)
     @FEATURE: #[stable(feature = "simd_x86", since = "1.27.0")] fma: "fma";

crates/std_detect/src/detect/os/x86.rs

@@ -209,6 +209,14 @@ pub(crate) fn detect_features() -> cache::Initializer {
                         enable(extended_features_ebx, 30, Feature::avx512bw);
                         enable(extended_features_ebx, 31, Feature::avx512vl);
                         enable(extended_features_ecx, 1, Feature::avx512vbmi);
+                        enable(extended_features_ecx, 5, Feature::avx512bf16);
+                        enable(extended_features_ecx, 6, Feature::avx512vbmi2);
+                        enable(extended_features_ecx, 8, Feature::avx512gfni);
+                        enable(extended_features_ecx, 8, Feature::avx512vp2intersect);
+                        enable(extended_features_ecx, 9, Feature::avx512vaes);
+                        enable(extended_features_ecx, 10, Feature::avx512vpclmulqdq);
+                        enable(extended_features_ecx, 11, Feature::avx512vnni);
+                        enable(extended_features_ecx, 12, Feature::avx512bitalg);
                         enable(extended_features_ecx, 14, Feature::avx512vpopcntdq);
                     }
                 }

crates/std_detect/tests/cpu-detection.rs

@@ -94,6 +94,23 @@ fn x86_all() {
         "avx512vpopcntdq: {:?}",
         is_x86_feature_detected!("avx512vpopcntdq")
     );
+    println!("avx512vbmi2 {:?}", is_x86_feature_detected!("avx512vbmi2"));
+    println!("avx512gfni {:?}", is_x86_feature_detected!("avx512gfni"));
+    println!("avx512vaes {:?}", is_x86_feature_detected!("avx512vaes"));
+    println!(
+        "avx512vpclmulqdq {:?}",
+        is_x86_feature_detected!("avx512vpclmulqdq")
+    );
+    println!("avx512vnni {:?}", is_x86_feature_detected!("avx512vnni"));
+    println!(
+        "avx512bitalg {:?}",
+        is_x86_feature_detected!("avx512bitalg")
+    );
+    println!("avx512bf16 {:?}", is_x86_feature_detected!("avx512bf16"));
+    println!(
+        "avx512vp2intersect {:?}",
+        is_x86_feature_detected!("avx512vp2intersect")
+    );
     println!("f16c: {:?}", is_x86_feature_detected!("f16c"));
     println!("fma: {:?}", is_x86_feature_detected!("fma"));
     println!("bmi1: {:?}", is_x86_feature_detected!("bmi1"));

crates/std_detect/tests/x86-specific.rs

@@ -35,6 +35,23 @@ fn dump() {
         "avx512_vpopcntdq {:?}",
         is_x86_feature_detected!("avx512vpopcntdq")
     );
+    println!("avx512vbmi2 {:?}", is_x86_feature_detected!("avx512vbmi2"));
+    println!("avx512gfni {:?}", is_x86_feature_detected!("avx512gfni"));
+    println!("avx512vaes {:?}", is_x86_feature_detected!("avx512vaes"));
+    println!(
+        "avx512vpclmulqdq {:?}",
+        is_x86_feature_detected!("avx512vpclmulqdq")
+    );
+    println!("avx512vnni {:?}", is_x86_feature_detected!("avx512vnni"));
+    println!(
+        "avx512bitalg {:?}",
+        is_x86_feature_detected!("avx512bitalg")
+    );
+    println!("avx512bf16 {:?}", is_x86_feature_detected!("avx512bf16"));
+    println!(
+        "avx512vp2intersect {:?}",
+        is_x86_feature_detected!("avx512vp2intersect")
+    );
     println!("fma: {:?}", is_x86_feature_detected!("fma"));
     println!("abm: {:?}", is_x86_feature_detected!("abm"));
     println!("bmi: {:?}", is_x86_feature_detected!("bmi1"));