ARM64 SHA256 Crypto intrinsics implementation

10 files changed, 117 insertions, 5 deletions

diff --git a/src/inc/corjitflags.h b/src/inc/corjitflags.h
index da303b6a03..84fb42f083 100644
--- a/src/inc/corjitflags.h
+++ b/src/inc/corjitflags.h
@@ -120,7 +120,7 @@ public:
         CORJIT_FLAG_HAS_ARM64_LRCPC         = 52, // ID_AA64ISAR1_EL1.LRCPC is 1 or better
         CORJIT_FLAG_HAS_ARM64_PMULL         = 53, // ID_AA64ISAR0_EL1.AES is 2 or better
         CORJIT_FLAG_HAS_ARM64_SHA1          = 54, // ID_AA64ISAR0_EL1.SHA1 is 1 or better
-        CORJIT_FLAG_HAS_ARM64_SHA2          = 55, // ID_AA64ISAR0_EL1.SHA2 is 1 or better
+        CORJIT_FLAG_HAS_ARM64_SHA256        = 55, // ID_AA64ISAR0_EL1.SHA2 is 1 or better
         CORJIT_FLAG_HAS_ARM64_SHA512        = 56, // ID_AA64ISAR0_EL1.SHA2 is 2 or better
         CORJIT_FLAG_HAS_ARM64_SHA3          = 57, // ID_AA64ISAR0_EL1.SHA3 is 1 or better
         CORJIT_FLAG_HAS_ARM64_SIMD          = 58, // ID_AA64PFR0_EL1.AdvSIMD is 0 or better
diff --git a/src/jit/emitarm64.cpp b/src/jit/emitarm64.cpp
index 1b41280523..7d5af2e006 100644
--- a/src/jit/emitarm64.cpp
+++ b/src/jit/emitarm64.cpp
@@ -4316,6 +4316,7 @@ void emitter::emitIns_R_R(
             fmt = IF_DR_2J;
             break;
 
+        case INS_sha256su0:
         case INS_sha1su1:
             assert(isVectorRegister(reg1));
             assert(isVectorRegister(reg2));
@@ -5387,6 +5388,9 @@ void emitter::emitIns_R_R_R(
             fmt = IF_LS_3D;
             break;
 
+        case INS_sha256h:
+        case INS_sha256h2:
+        case INS_sha256su1:
         case INS_sha1su0:
         case INS_sha1c:
         case INS_sha1p:
@@ -11438,6 +11442,13 @@ void emitter::emitDispIns(
                 emitDispReg(id->idReg2(), EA_4BYTE, true);
                 emitDispVectorReg(id->idReg3(), id->idInsOpt(), false);
             }
+            else if ((ins == INS_sha256h) || (ins == INS_sha256h2))
+            {
+                // Qd Qn Vm (vector)
+                emitDispReg(id->idReg1(), size, true);
+                emitDispReg(id->idReg2(), size, true);
+                emitDispVectorReg(id->idReg3(), id->idInsOpt(), false);
+            }
             else
             {
                 emitDispVectorReg(id->idReg1(), id->idInsOpt(), true);
diff --git a/src/jit/hwintrinsiclistArm64.h b/src/jit/hwintrinsiclistArm64.h
index 05d910f51a..ffc8d39c52 100644
--- a/src/jit/hwintrinsiclistArm64.h
+++ b/src/jit/hwintrinsiclistArm64.h
@@ -23,7 +23,7 @@ HARDWARE_INTRINSIC_CLASS(JIT_FLAG_HAS_ARM64_JSCVT     , Jscvt    )
 HARDWARE_INTRINSIC_CLASS(JIT_FLAG_HAS_ARM64_LRCPC     , Lrcpc    )
 HARDWARE_INTRINSIC_CLASS(JIT_FLAG_HAS_ARM64_PMULL     , Pmull    )
 HARDWARE_INTRINSIC_CLASS(JIT_FLAG_HAS_ARM64_SHA1      , Sha1     )
-HARDWARE_INTRINSIC_CLASS(JIT_FLAG_HAS_ARM64_SHA2      , Sha2     )
+HARDWARE_INTRINSIC_CLASS(JIT_FLAG_HAS_ARM64_SHA256    , Sha256   )
 HARDWARE_INTRINSIC_CLASS(JIT_FLAG_HAS_ARM64_SHA512    , Sha512   )
 HARDWARE_INTRINSIC_CLASS(JIT_FLAG_HAS_ARM64_SHA3      , Sha3     )
 HARDWARE_INTRINSIC_CLASS(JIT_FLAG_HAS_ARM64_SIMD      , Simd     )
@@ -94,6 +94,11 @@ HARDWARE_INTRINSIC(NI_ARM64_Sha1FixedRotate,     Sha1,      FixedRotate,
 HARDWARE_INTRINSIC(NI_ARM64_Sha1SchedulePart1,   Sha1,      SchedulePart1,                 SimdTernaryRMWOp,       INS_invalid,    INS_invalid, INS_sha1su0,    None )
 HARDWARE_INTRINSIC(NI_ARM64_Sha1SchedulePart2,   Sha1,      SchedulePart2,                 SimdBinaryRMWOp,        INS_invalid,    INS_invalid, INS_sha1su1,    None )
 
+//Sha256
+HARDWARE_INTRINSIC(NI_ARM64_Sha256HashLower,       Sha256,      HashLower,                 SimdTernaryRMWOp,   INS_invalid,    INS_invalid, INS_sha256h,    None )
+HARDWARE_INTRINSIC(NI_ARM64_Sha256HashUpper,       Sha256,      HashUpper,                 SimdTernaryRMWOp,   INS_invalid,    INS_invalid, INS_sha256h2,   None )
+HARDWARE_INTRINSIC(NI_ARM64_Sha256SchedulePart1,   Sha256,      SchedulePart1,             SimdBinaryRMWOp,    INS_invalid,    INS_invalid, INS_sha256su0,  None )
+HARDWARE_INTRINSIC(NI_ARM64_Sha256SchedulePart2,   Sha256,      SchedulePart2,             SimdTernaryRMWOp,   INS_invalid,    INS_invalid, INS_sha256su1,  None )
 #endif
 
 
diff --git a/src/jit/instr.h b/src/jit/instr.h
index 7838fa44e4..c7be413d05 100644
--- a/src/jit/instr.h
+++ b/src/jit/instr.h
@@ -314,7 +314,7 @@ enum InstructionSet
     InstructionSet_Lrcpc,     // ID_AA64ISAR1_EL1.LRCPC is 1 or better
     InstructionSet_Pmull,     // ID_AA64ISAR0_EL1.AES is 2 or better
     InstructionSet_Sha1,      // ID_AA64ISAR0_EL1.SHA1 is 1 or better
-    InstructionSet_Sha2,      // ID_AA64ISAR0_EL1.SHA2 is 1 or better
+    InstructionSet_Sha256,    // ID_AA64ISAR0_EL1.SHA2 is 1 or better
     InstructionSet_Sha512,    // ID_AA64ISAR0_EL1.SHA2 is 2 or better
     InstructionSet_Sha3,      // ID_AA64ISAR0_EL1.SHA3 is 1 or better
     InstructionSet_Simd,      // ID_AA64PFR0_EL1.AdvSIMD is 0 or better
diff --git a/src/jit/instrsarm64.h b/src/jit/instrsarm64.h
index ee5feae576..8f5971bcb7 100644
--- a/src/jit/instrsarm64.h
+++ b/src/jit/instrsarm64.h
@@ -935,6 +935,18 @@ INST1(sha1su0, "sha1su0",  0, 0, IF_DV_3F,  0x5E003000)
 
 INST1(sha1su1, "sha1su1",  0, 0, IF_DV_2P,  0x5E281800)
                                    //  sha1su1 Vd.4S, Vn.4S         DV_2P  0101111000101000 000110nnnnnddddd   5E28 1800   Vd.4S Vn.4S   (vector)
+
+INST1(sha256h, "sha256h",  0, 0, IF_DV_3F,  0x5E004000)
+                                   //  sha256h  Qd,Qn,Vm.4S         DV_3F  01011110000mmmmm 010000nnnnnddddd   5E00 4000  Qd Qn Vm.4S   (vector)
+
+INST1(sha256h2, "sha256h2",  0, 0, IF_DV_3F,  0x5E005000)
+                                   //  sha256h  Qd,Qn,Vm.4S         DV_3F  01011110000mmmmm 010100nnnnnddddd   5E00 5000  Qd Qn Vm.4S   (vector)
+
+INST1(sha256su0, "sha256su0",  0, 0, IF_DV_2P,  0x5E282800)
+                                   // sha256su0  Vd.4S,Vn.4S        DV_2P  0101111000101000 001010nnnnnddddd   5E28 2800  Vd.4S Vn.4S   (vector)
+
+INST1(sha256su1, "sha256su1",  0, 0, IF_DV_3F,  0x5E006000)
+                                   // sha256su1  Vd.4S,Vn.4S,Vm.4S  DV_3F  01011110000mmmmm 011000nnnnnddddd   5E00 6000  Vd.4S Vn.4S Vm.4S   (vector)
    
 INST1(sbfm,    "sbfm",   0, 0, IF_DI_2D,  0x13000000)
                                    //  sbfm    Rd,Rn,imr,ims        DI_2D  X00100110Nrrrrrr ssssssnnnnnddddd   1300 0000   imr, ims
diff --git a/src/jit/jitee.h b/src/jit/jitee.h
index b36e7661fd..903b3cffd9 100644
--- a/src/jit/jitee.h
+++ b/src/jit/jitee.h
@@ -107,7 +107,7 @@ public:
         JIT_FLAG_HAS_ARM64_LRCPC         = 52, // ID_AA64ISAR1_EL1.LRCPC is 1 or better
         JIT_FLAG_HAS_ARM64_PMULL         = 53, // ID_AA64ISAR0_EL1.AES is 2 or better
         JIT_FLAG_HAS_ARM64_SHA1          = 54, // ID_AA64ISAR0_EL1.SHA1 is 1 or better
-        JIT_FLAG_HAS_ARM64_SHA2          = 55, // ID_AA64ISAR0_EL1.SHA2 is 1 or better
+        JIT_FLAG_HAS_ARM64_SHA256        = 55, // ID_AA64ISAR0_EL1.SHA2 is 1 or better
         JIT_FLAG_HAS_ARM64_SHA512        = 56, // ID_AA64ISAR0_EL1.SHA2 is 2 or better
         JIT_FLAG_HAS_ARM64_SHA3          = 57, // ID_AA64ISAR0_EL1.SHA3 is 1 or better
         JIT_FLAG_HAS_ARM64_SIMD          = 58, // ID_AA64PFR0_EL1.AdvSIMD is 0 or better
diff --git a/src/mscorlib/System.Private.CoreLib.csproj b/src/mscorlib/System.Private.CoreLib.csproj
index 44a432f2ef..1da5ccba25 100644
--- a/src/mscorlib/System.Private.CoreLib.csproj
+++ b/src/mscorlib/System.Private.CoreLib.csproj
@@ -300,11 +300,13 @@
     <Compile Include="$(BclSourcesRoot)\System\Runtime\Intrinsics\Arm\Arm64\Simd.cs" />
     <Compile Include="$(BclSourcesRoot)\System\Runtime\Intrinsics\Arm\Arm64\Aes.cs" />
     <Compile Include="$(BclSourcesRoot)\System\Runtime\Intrinsics\Arm\Arm64\Sha1.cs" />
+    <Compile Include="$(BclSourcesRoot)\System\Runtime\Intrinsics\Arm\Arm64\Sha256.cs" />
   </ItemGroup>
   <ItemGroup Condition="'$(Platform)' != 'arm64'">
     <Compile Include="$(BclSourcesRoot)\System\Runtime\Intrinsics\Arm\Arm64\Simd.PlatformNotSupported.cs" />
     <Compile Include="$(BclSourcesRoot)\System\Runtime\Intrinsics\Arm\Arm64\Aes.PlatformNotSupported.cs" />
     <Compile Include="$(BclSourcesRoot)\System\Runtime\Intrinsics\Arm\Arm64\Sha1.PlatformNotSupported.cs" />
+    <Compile Include="$(BclSourcesRoot)\System\Runtime\Intrinsics\Arm\Arm64\Sha256.PlatformNotSupported.cs" />
   </ItemGroup>
   <ItemGroup>
     <Compile Include="$(BclSourcesRoot)\System\AppContext\AppContext.cs" />
diff --git a/src/mscorlib/src/System/Runtime/Intrinsics/Arm/Arm64/Sha256.PlatformNotSupported.cs b/src/mscorlib/src/System/Runtime/Intrinsics/Arm/Arm64/Sha256.PlatformNotSupported.cs
new file mode 100644
index 0000000000..052a1d236b
--- /dev/null
+++ b/src/mscorlib/src/System/Runtime/Intrinsics/Arm/Arm64/Sha256.PlatformNotSupported.cs
@@ -0,0 +1,41 @@
+using System.Runtime.CompilerServices;
+using System.Runtime.Intrinsics;
+
+namespace System.Runtime.Intrinsics.Arm.Arm64
+{
+    /// <summary>
+    /// This class provides access to the Arm64 SHA256 Crypto intrinsics
+    ///
+    /// Arm64 CPU indicate support for this feature by setting
+    /// ID_AA64ISAR0_EL1.SHA2 is 1 or better
+    /// </summary>
+    [CLSCompliant(false)]
+    public static class Sha256
+    {
+        public static bool IsSupported { get { return false; } }
+
+        // <summary>
+        /// Performs SHA256 hash update (part 1).
+        /// vsha256hq_u32 (uint32x4_t hash_abcd, uint32x4_t hash_efgh, uint32x4_t wk)
+        ///</summary>
+        public static Vector128<uint> HashLower(Vector128<uint> hash_abcd, Vector128<uint> hash_efgh, Vector128<uint> wk) { throw new PlatformNotSupportedException(); }
+
+        // <summary>
+        /// Performs SHA256 hash update (part 2).
+        /// vsha256h2q_u32 (uint32x4_t hash_efgh, uint32x4_t hash_abcd, uint32x4_t wk)
+        ///</summary>
+        public static Vector128<uint> HashUpper(Vector128<uint> hash_efgh, Vector128<uint> hash_abcd, Vector128<uint> wk) { throw new PlatformNotSupportedException(); }
+
+        // <summary>
+        /// Performs SHA256 schedule update 0
+        /// vsha256su0q_u32 (uint32x4_t w0_3, uint32x4_t w4_7)
+        ///</summary>
+        public static Vector128<uint> SchedulePart1(Vector128<uint> w0_3, Vector128<uint> w4_7) { throw new PlatformNotSupportedException(); }
+
+        // <summary>
+        /// Performs SHA256 schedule update 1
+        /// vsha256su1q_u32 (uint32x4_t w0_3, uint32x4_t w8_11, uint32x4_t w12_15)
+        ///</summary>
+        public static Vector128<uint> SchedulePart2(Vector128<uint> w0_3, Vector128<uint> w8_11, Vector128<uint> w12_15) { throw new PlatformNotSupportedException(); }
+    }
+}
diff --git a/src/mscorlib/src/System/Runtime/Intrinsics/Arm/Arm64/Sha256.cs b/src/mscorlib/src/System/Runtime/Intrinsics/Arm/Arm64/Sha256.cs
new file mode 100644
index 0000000000..4e7e51db30
--- /dev/null
+++ b/src/mscorlib/src/System/Runtime/Intrinsics/Arm/Arm64/Sha256.cs
@@ -0,0 +1,41 @@
+using System.Runtime.CompilerServices;
+using System.Runtime.Intrinsics;
+
+namespace System.Runtime.Intrinsics.Arm.Arm64
+{
+    /// <summary>
+    /// This class provides access to the Arm64 SHA256 Crypto intrinsics
+    ///
+    /// Arm64 CPU indicate support for this feature by setting
+    /// ID_AA64ISAR0_EL1.SHA2 is 1 or better
+    /// </summary>
+    [CLSCompliant(false)]
+    public static class Sha256
+    {
+        public static bool IsSupported { get => IsSupported; }
+
+        // <summary>
+        /// Performs SHA256 hash update (part 1).
+        /// vsha256hq_u32 (uint32x4_t hash_abcd, uint32x4_t hash_efgh, uint32x4_t wk)
+        ///</summary>
+        public static Vector128<uint> HashLower(Vector128<uint> hash_abcd, Vector128<uint> hash_efgh, Vector128<uint> wk) => HashLower(hash_abcd, hash_efgh, wk);
+
+        // <summary>
+        /// Performs SHA256 hash update (part 2).
+        /// vsha256h2q_u32 (uint32x4_t hash_efgh, uint32x4_t hash_abcd, uint32x4_t wk)
+        ///</summary>
+        public static Vector128<uint> HashUpper(Vector128<uint> hash_efgh, Vector128<uint> hash_abcd, Vector128<uint> wk) => HashUpper(hash_efgh, hash_abcd, wk);
+
+        // <summary>
+        /// Performs SHA256 schedule update 0
+        /// vsha256su0q_u32 (uint32x4_t w0_3, uint32x4_t w4_7)
+        ///</summary>
+        public static Vector128<uint> SchedulePart1(Vector128<uint> w0_3, Vector128<uint> w4_7) => SchedulePart1(w0_3, w4_7);
+
+        // <summary>
+        /// Performs SHA256 schedule update 1
+        /// vsha256su1q_u32 (uint32x4_t tw0_3, uint32x4_t w8_11, uint32x4_t w12_15)
+        ///</summary>
+        public static Vector128<uint> SchedulePart2(Vector128<uint> w0_3, Vector128<uint> w8_11, Vector128<uint> w12_15) => SchedulePart2(w0_3, w8_11, w12_15);
+    }
+}
diff --git a/src/pal/src/misc/jitsupport.cpp b/src/pal/src/misc/jitsupport.cpp
index 917319d7c5..d3d4f1e5c5 100644
--- a/src/pal/src/misc/jitsupport.cpp
+++ b/src/pal/src/misc/jitsupport.cpp
@@ -86,7 +86,7 @@ PAL_GetJitCpuCapabilityFlags(CORJIT_FLAGS *flags)
 #endif
 #ifdef HWCAP_SHA2
     if (hwCap & HWCAP_SHA2)
-        CPUCompileFlags.Set(CORJIT_FLAGS::CORJIT_FLAG_HAS_ARM64_SHA2);
+        CPUCompileFlags.Set(CORJIT_FLAGS::CORJIT_FLAG_HAS_ARM64_SHA256);
 #endif
 #ifdef HWCAP_SHA512
     if (hwCap & HWCAP_SHA512)