Add feature detection for new amx variants and movrs

2 files changed, 42 insertions, 9 deletions

diff --git a/library/stdarch/crates/std_detect/src/detect/arch/x86.rs b/library/stdarch/crates/std_detect/src/detect/arch/x86.rs
index 6c45c76e9db..7fdb72bb4ae 100644
--- a/library/stdarch/crates/std_detect/src/detect/arch/x86.rs
+++ b/library/stdarch/crates/std_detect/src/detect/arch/x86.rs
@@ -89,6 +89,11 @@ features! {
     /// * `"amx-bf16"`
     /// * `"amx-fp16"`
     /// * `"amx-complex"`
+    /// * `"amx-avx512"`
+    /// * `"amx-fp8"`
+    /// * `"amx-movrs"`
+    /// * `"amx-tf32"`
+    /// * `"amx-transpose"`
     /// * `"f16c"`
     /// * `"fma"`
     /// * `"bmi1"`
@@ -109,6 +114,8 @@ features! {
     /// * `"rtm"`
     /// * `"movbe"`
     /// * `"ermsb"`
+    /// * `"movrs"`
+    /// * `"xop"`
     ///
     /// [docs]: https://software.intel.com/sites/landingpage/IntrinsicsGuide
     #[stable(feature = "simd_x86", since = "1.27.0")]
@@ -177,8 +184,7 @@ features! {
     @FEATURE: #[stable(feature = "simd_x86", since = "1.27.0")] avx512vbmi: "avx512vbmi";
     /// AVX-512 VBMI (Vector Byte Manipulation Instructions)
     @FEATURE: #[stable(feature = "simd_x86", since = "1.27.0")] avx512vpopcntdq: "avx512vpopcntdq";
-    /// AVX-512 VPOPCNTDQ (Vector Population Count Doubleword and
-    /// Quadword)
+    /// AVX-512 VPOPCNTDQ (Vector Population Count Doubleword and Quadword)
     @FEATURE: #[stable(feature = "simd_x86", since = "1.27.0")] avx512vbmi2: "avx512vbmi2";
     /// AVX-512 VBMI2 (Additional byte, word, dword and qword capabilities)
     @FEATURE: #[stable(feature = "simd_x86", since = "1.27.0")] gfni: "gfni";
@@ -217,6 +223,16 @@ features! {
     /// AMX-FP16 (Float16 Operations)
     @FEATURE: #[unstable(feature = "x86_amx_intrinsics", issue = "126622")] amx_complex: "amx-complex";
     /// AMX-COMPLEX (Complex number Operations)
+    @FEATURE: #[unstable(feature = "x86_amx_intrinsics", issue = "126622")] amx_avx512: "amx-avx512";
+    /// AMX-AVX512 (AVX512 operations extended to matrices)
+    @FEATURE: #[unstable(feature = "x86_amx_intrinsics", issue = "126622")] amx_fp8: "amx-fp8";
+    /// AMX-FP8 (Float8 Operations)
+    @FEATURE: #[unstable(feature = "x86_amx_intrinsics", issue = "126622")] amx_movrs: "amx-movrs";
+    /// AMX-MOVRS (Matrix MOVERS operations)
+    @FEATURE: #[unstable(feature = "x86_amx_intrinsics", issue = "126622")] amx_tf32: "amx-tf32";
+    /// AMX-TF32 (TensorFloat32 Operations)
+    @FEATURE: #[unstable(feature = "x86_amx_intrinsics", issue = "126622")] amx_transpose: "amx-transpose";
+    /// AMX-TRANSPOSE (Matrix Transpose Operations)
     @FEATURE: #[stable(feature = "simd_x86", since = "1.27.0")] f16c: "f16c";
     /// F16C (Conversions between IEEE-754 `binary16` and `binary32` formats)
     @FEATURE: #[stable(feature = "simd_x86", since = "1.27.0")] fma: "fma";
@@ -253,6 +269,8 @@ features! {
     /// RTM, Intel (Restricted Transactional Memory)
     @FEATURE: #[stable(feature = "movbe_target_feature", since = "1.67.0")] movbe: "movbe";
     /// MOVBE (Move Data After Swapping Bytes)
+    @FEATURE: #[unstable(feature = "movrs_target_feature", issue = "137976")] movrs: "movrs";
+    /// MOVRS (Move data with the read-shared hint)
     @FEATURE: #[stable(feature = "simd_x86", since = "1.27.0")] ermsb: "ermsb";
     /// ERMSB, Enhanced REP MOVSB and STOSB
     @FEATURE: #[unstable(feature = "xop_target_feature", issue = "127208")] xop: "xop";
diff --git a/library/stdarch/crates/std_detect/src/detect/os/x86.rs b/library/stdarch/crates/std_detect/src/detect/os/x86.rs
index bb6a44b6438..e48d04ad004 100644
--- a/library/stdarch/crates/std_detect/src/detect/os/x86.rs
+++ b/library/stdarch/crates/std_detect/src/detect/os/x86.rs
@@ -141,6 +141,8 @@ pub(crate) fn detect_features() -> cache::Initializer {
 
         enable(extended_features_ebx, 9, Feature::ermsb);
 
+        enable(extended_features_eax_leaf_1, 31, Feature::movrs);
+
         // Detect if CPUID.19h available
         if bit::test(extended_features_ecx as usize, 23) {
             let CpuidResult { ebx, .. } = unsafe { __cpuid(0x19) };
@@ -250,14 +252,27 @@ pub(crate) fn detect_features() -> cache::Initializer {
                         enable(extended_features_edx, 8, Feature::avx512vp2intersect);
                         enable(extended_features_edx, 23, Feature::avx512fp16);
                         enable(extended_features_eax_leaf_1, 5, Feature::avx512bf16);
+                    }
+                }
+
+                if os_amx_support {
+                    enable(extended_features_edx, 24, Feature::amx_tile);
+                    enable(extended_features_edx, 25, Feature::amx_int8);
+                    enable(extended_features_edx, 22, Feature::amx_bf16);
+                    enable(extended_features_eax_leaf_1, 21, Feature::amx_fp16);
+                    enable(extended_features_edx_leaf_1, 8, Feature::amx_complex);
+
+                    if max_basic_leaf >= 0x1e {
+                        let CpuidResult {
+                            eax: amx_feature_flags_eax,
+                            ..
+                        } = unsafe { __cpuid_count(0x1e_u32, 1) };
 
-                        if os_amx_support {
-                            enable(extended_features_edx, 24, Feature::amx_tile);
-                            enable(extended_features_edx, 25, Feature::amx_int8);
-                            enable(extended_features_edx, 22, Feature::amx_bf16);
-                            enable(extended_features_eax_leaf_1, 21, Feature::amx_fp16);
-                            enable(extended_features_edx_leaf_1, 8, Feature::amx_complex);
-                        }
+                        enable(amx_feature_flags_eax, 4, Feature::amx_fp8);
+                        enable(amx_feature_flags_eax, 5, Feature::amx_transpose);
+                        enable(amx_feature_flags_eax, 6, Feature::amx_tf32);
+                        enable(amx_feature_flags_eax, 7, Feature::amx_avx512);
+                        enable(amx_feature_flags_eax, 8, Feature::amx_movrs);
                     }
                 }
             }