New X86 features and detections

Author: yuyichao

src/features_x86.h

@@ -56,11 +56,31 @@ JL_FEATURE_DEF(avx512vl, 32 * 2 + 31, 0)
 JL_FEATURE_DEF(prefetchwt1, 32 * 3 + 0, 0)
 JL_FEATURE_DEF(avx512vbmi, 32 * 3 + 1, 0)
 JL_FEATURE_DEF(pku, 32 * 3 + 4, 0) // ospke
+JL_FEATURE_DEF(waitpkg, 32 * 3 + 5, 0)
+JL_FEATURE_DEF(avx512vbmi2, 32 * 3 + 6, 0)
+JL_FEATURE_DEF(shstk, 32 * 3 + 7, 0)
+JL_FEATURE_DEF(gfni, 32 * 3 + 8, 0)
+JL_FEATURE_DEF(vaes, 32 * 3 + 9, 0)
+JL_FEATURE_DEF(vpclmulqdq, 32 * 3 + 10, 0)
+JL_FEATURE_DEF(avx512vnni, 32 * 3 + 11, 0)
+JL_FEATURE_DEF(avx512bitalg, 32 * 3 + 12, 0)
 JL_FEATURE_DEF(avx512vpopcntdq, 32 * 3 + 14, 0)
+JL_FEATURE_DEF(rdpid, 32 * 3 + 22, 0)
+JL_FEATURE_DEF(cldemote, 32 * 3 + 25, 0)
+JL_FEATURE_DEF(movdiri, 32 * 3 + 27, 0)
+JL_FEATURE_DEF(movdir64b, 32 * 3 + 28, 0)
+JL_FEATURE_DEF(enqcmd, 32 * 3 + 29, 90000)
 
 // EAX=7,ECX=0: EDX
 // JL_FEATURE_DEF(avx5124vnniw, 32 * 4 + 2, ?????)
 // JL_FEATURE_DEF(avx5124fmaps, 32 * 4 + 3, ?????)
+JL_FEATURE_DEF(avx512vp2intersect, 32 * 4 + 8, 90000)
+JL_FEATURE_DEF(serialize, 32 * 4 + 14, 110000)
+JL_FEATURE_DEF(tsxldtrk, 32 * 4 + 16, 110000)
+JL_FEATURE_DEF(pconfig, 32 * 4 + 18, 0)
+JL_FEATURE_DEF_NAME(amx_bf16, 32 * 4 + 22, 110000, "amx-bf16")
+JL_FEATURE_DEF_NAME(amx_tile, 32 * 4 + 24, 110000, "amx-tile")
+JL_FEATURE_DEF_NAME(amx_int8, 32 * 4 + 25, 110000, "amx-int8")
 
 // EAX=0x80000001: ECX
 // ignore sahf on 32bit x86 since it is required
@@ -85,5 +105,12 @@ JL_FEATURE_DEF(xsaves, 32 * 7 + 3, 0)
 
 // EAX=0x80000008: EBX
 JL_FEATURE_DEF(clzero, 32 * 8 + 0, 0)
+JL_FEATURE_DEF(wbnoinvd, 32 * 8 + 9, 0)
+
+// EAX=7,ECX=1: EAX
+JL_FEATURE_DEF(avx512bf16, 32 * 9 + 5, 90000)
+
+// EAX=0x14,ECX=0: EBX
+JL_FEATURE_DEF(ptwrite, 32 * 10 + 4, 0)
 
 #undef JL_X86_64ONLY_VER

src/processor_x86.cpp

@@ -92,7 +92,7 @@ enum class CPU : uint32_t {
     amd_znver1,
 };
 
-static constexpr size_t feature_sz = 9;
+static constexpr size_t feature_sz = 11;
 static constexpr FeatureName feature_names[] = {
 #define JL_FEATURE_DEF(name, bit, llvmver) {#name, bit, llvmver},
 #define JL_FEATURE_DEF_NAME(name, bit, llvmver, str) {str, bit, llvmver},
@@ -130,16 +130,27 @@ static constexpr FeatureDep deps[] = {
     {avx, sse42},
     {f16c, avx},
     {avx2, avx},
+    {vaes, avx},
+    {vaes, aes},
+    {vpclmulqdq, avx},
+    {vpclmulqdq, pclmul},
     {avx512f, avx2},
     {avx512dq, avx512f},
     {avx512ifma, avx512f},
     {avx512pf, avx512f},
     {avx512er, avx512f},
     {avx512cd, avx512f},
     {avx512bw, avx512f},
+    {avx512bf16, avx512bw},
+    {avx512bitalg, avx512bw},
     {avx512vl, avx512f},
     {avx512vbmi, avx512bw},
+    {avx512vbmi2, avx512bw},
+    {avx512vnni, avx512f},
+    {avx512vp2intersect, avx512f},
     {avx512vpopcntdq, avx512f},
+    {amx_int8, amx_tile},
+    {amx_bf16, amx_tile},
     {sse4a, sse3},
     {xop, fma4},
     {fma4, avx},
@@ -470,15 +481,23 @@ static inline void features_disable_avx512(T &features)
 {
     using namespace Feature;
     unset_bits(features, avx512f, avx512dq, avx512ifma, avx512pf, avx512er, avx512cd,
-               avx512bw, avx512vl, avx512vbmi);
+               avx512bw, avx512vl, avx512vbmi, avx512vpopcntdq, avx512vbmi2, avx512vnni,
+               avx512bitalg, avx512vp2intersect, avx512bf16);
 }
 
 template<typename T>
 static inline void features_disable_avx(T &features)
 {
     using namespace Feature;
     unset_bits(features, avx, Feature::fma, f16c, xsave, avx2, xop, fma4,
-               xsaveopt, xsavec, xsaves);
+               xsaveopt, xsavec, xsaves, vaes, vpclmulqdq);
+}
+
+template<typename T>
+static inline void features_disable_amx(T &features)
+{
+    using namespace Feature;
+    unset_bits(features, amx_bf16, amx_tile, amx_int8);
 }
 
 static NOINLINE std::pair<uint32_t,FeatureList<feature_sz>> _get_host_cpu(void)
@@ -535,15 +554,25 @@ static NOINLINE std::pair<uint32_t,FeatureList<feature_sz>> _get_host_cpu(void)
         jl_cpuidex(infoex8, 0x80000008, 0);
         features[8] = infoex8[1];
     }
+    if (maxleaf >= 7) {
+        int32_t info7[4];
+        jl_cpuidex(info7, 7, 1);
+        features[9] = info7[0];
+    }
+    if (maxleaf >= 0x14) {
+        int32_t info14[4];
+        jl_cpuidex(info14, 0x14, 0);
+        features[10] = info14[1];
+    }
 
     // Fix up AVX bits to account for OS support and match LLVM model
     uint64_t xcr0 = 0;
-    const uint32_t avx_mask = (1 << 27) | (1 << 28);
-    bool hasavx = test_all_bits(features[0], avx_mask);
-    if (hasavx) {
+    bool hasxsave = test_all_bits(features[0], 1 << 27);
+    if (hasxsave) {
         xcr0 = get_xcr0();
-        hasavx = test_all_bits(xcr0, 0x6);
+        hasxsave = test_all_bits(xcr0, 0x6);
     }
+    bool hasavx = hasxsave && test_all_bits(features[0], 1 << 28);
     unset_bits(features, 32 + 27);
     if (!hasavx)
         features_disable_avx(features);
@@ -557,6 +586,10 @@ static NOINLINE std::pair<uint32_t,FeatureList<feature_sz>> _get_host_cpu(void)
 #endif
     if (!hasavx512save)
         features_disable_avx512(features);
+    // AMX requires additional context to be saved by the OS.
+    bool hasamxsave = hasxsave && test_all_bits(xcr0, (1 << 17) | (1 << 18));
+    if (!hasamxsave)
+        features_disable_amx(features);
     // Ignore feature bits that we are not interested in.
     mask_features(feature_masks, &features[0]);
 
@@ -788,12 +821,16 @@ static void ensure_jit_target(bool imaging)
         static constexpr uint32_t clone_simd[] = {Feature::sse3, Feature::ssse3,
                                                   Feature::sse41, Feature::sse42,
                                                   Feature::avx, Feature::avx2,
+                                                  Feature::vaes, Feature::vpclmulqdq,
                                                   Feature::sse4a, Feature::avx512f,
                                                   Feature::avx512dq, Feature::avx512ifma,
                                                   Feature::avx512pf, Feature::avx512er,
                                                   Feature::avx512cd, Feature::avx512bw,
                                                   Feature::avx512vl, Feature::avx512vbmi,
-                                                  Feature::avx512vpopcntdq};
+                                                  Feature::avx512vpopcntdq,
+                                                  Feature::avx512vbmi2, Feature::avx512vnni,
+                                                  Feature::avx512bitalg, Feature::avx512bf16,
+                                                  Feature::avx512vp2intersect};
         for (auto fe: clone_math) {
             if (!test_nbit(features0, fe) && test_nbit(t.en.features, fe)) {
                 t.en.flags |= JL_TARGET_CLONE_MATH;
@@ -847,6 +884,9 @@ get_llvm_target_noext(const TargetData<feature_sz> &data)
     // returns a value that may not have 64bit support.
     // This can happen with virtualization.
     features.push_back("+64bit");
+#endif
+#if JL_LLVM_VERSION >= 90000
+    features.push_back("+cx8");
 #endif
     return std::make_pair(std::move(name), std::move(features));
 }