Add BMI2 3DES implementation

2 files changed, 20 insertions, 0 deletions

diff --git a/src/lib/utils/cpuid/cpuid.h b/src/lib/utils/cpuid/cpuid.h
index db0bfe1a9..102ed9593 100644
--- a/src/lib/utils/cpuid/cpuid.h
+++ b/src/lib/utils/cpuid/cpuid.h
@@ -115,6 +115,7 @@ class BOTAN_PUBLIC_API(2,1) CPUID final
          CPUID_ADX_BIT        = (1ULL << 49),
          CPUID_BMI1_BIT       = (1ULL << 50),
          CPUID_BMI2_BIT       = (1ULL << 51),
+         CPUID_FAST_PDEP_BIT  = (1ULL << 52),
 #endif
 
 #if defined(BOTAN_TARGET_CPU_IS_PPC_FAMILY)
@@ -310,6 +311,12 @@ class BOTAN_PUBLIC_API(2,1) CPUID final
          { return has_cpuid_bit(CPUID_BMI2_BIT); }
 
       /**
+      * Check if the processor supports fast PDEP/PEXT from BMI2
+      */
+      static bool has_fast_pdep()
+         { return has_cpuid_bit(CPUID_FAST_PDEP_BIT); }
+
+      /**
       * Check if the processor supports AES-NI
       */
       static bool has_aes_ni()
diff --git a/src/lib/utils/cpuid/cpuid_x86.cpp b/src/lib/utils/cpuid/cpuid_x86.cpp
index b8ae54f87..51e51fa8e 100644
--- a/src/lib/utils/cpuid/cpuid_x86.cpp
+++ b/src/lib/utils/cpuid/cpuid_x86.cpp
@@ -152,7 +152,20 @@ uint64_t CPUID::CPUID_Data::detect_cpu_features(size_t* cache_line_size)
          implements BMI2 but not BMI1.
          */
          if(flags7 & x86_CPUID_7_bits::BMI2)
+            {
             features_detected |= CPUID::CPUID_BMI2_BIT;
+
+            /*
+            Up until Zen3, AMD CPUs with BMI2 support had microcoded
+            pdep/pext, which works but is very slow.
+
+            TODO: check for Zen3 here
+            */
+            if(is_intel)
+               {
+               features_detected |= CPUID::CPUID_FAST_PDEP_BIT;
+               }
+            }
          }
 
       if(flags7 & x86_CPUID_7_bits::AVX512_F)