Crypto: Make use of XAR in SHA1NEXTE when available

Lets us shave an instruction off on hardware that supports XAR.

Closes #5730

FEXCore/Source/Interface/Core/JIT/VectorOps.cpp

@@ -4655,6 +4655,31 @@ DEF_OP(VBlendImm) {
   }
 }
 
+DEF_OP(VXar) {
+  LOGMAN_THROW_A_FMT(HostSupportsSVE128 || HostSupportsSVE256, "Host must support SVE to use {}", __func__);
+
+  auto Op = IROp->C<IR::IROp_VXar>();
+  const auto SubRegSize = ConvertSubRegSize8(IROp);
+  const auto ElementSizeBits = IR::OpSizeAsBits(IROp->ElementSize);
+
+  const auto Dst = GetVReg(Node);
+  const auto LHS = GetVReg(Op->LHS);
+  const auto RHS = GetVReg(Op->RHS);
+  const auto Rotate = Op->Rotate;
+  LOGMAN_THROW_A_FMT(Rotate >= 1 && Rotate <= ElementSizeBits, "Rotate immediate must be within [1, {}]", ElementSizeBits);
+
+  if (Dst == LHS) {
+    xar(SubRegSize, Dst.Z(), RHS.Z(), Rotate);
+  } else if (Dst == RHS) {
+    movprfx(VTMP1.Z(), LHS.Z());
+    xar(SubRegSize, VTMP1.Z(), RHS.Z(), Rotate);
+    mov(Dst.Z(), VTMP1.Z());
+  } else {
+    movprfx(Dst.Z(), LHS.Z());
+    xar(SubRegSize, Dst.Z(), RHS.Z(), Rotate);
+  }
+}
+
 DEF_OP(VFCopySign) {
   auto Op = IROp->C<IR::IROp_VFCopySign>();
   const auto OpSize = IROp->Size;

FEXCore/Source/Interface/Core/OpcodeDispatcher/Crypto.cpp

@@ -26,15 +26,23 @@ void OpDispatchBuilder::SHA1NEXTEOp(OpcodeArgs) {
   Ref Dest = LoadSourceFPR(Op, Op->Dest, Op->Flags);
   Ref Src = LoadSourceFPR(Op, Op->Src[0], Op->Flags);
 
-  // ARMv8 SHA1 extension provides a `SHA1H` instruction which does a fixed rotate by 30.
-  // This only operates on element 0 rather than element 3. We don't have the luxury of rewriting the x86 SHA algorithm to take advantage of this.
-  // Move the element to zero, rotate, and then move back (Using duplicates).
-  // Saves one instruction versus that path that doesn't support SHA extension.
-  auto Duplicated = _VDupElement(OpSize::i128Bit, OpSize::i32Bit, Dest, 3);
-  auto Sha1HRotated = _VSha1H(Duplicated);
-  auto RotatedNode = _VDupElement(OpSize::i128Bit, OpSize::i32Bit, Sha1HRotated, 0);
-  auto Tmp = _VAdd(OpSize::i128Bit, OpSize::i32Bit, Src, RotatedNode);
-  auto Result = _VInsElement(OpSize::i128Bit, OpSize::i32Bit, 3, 3, Src, Tmp);
+  Ref Result {};
+  if (CTX->HostFeatures.SupportsSVE128) {
+    auto ZeroVec = LoadZeroVector(OpSize::i128Bit);
+    auto Tmp = _VInsElement(OpSize::i128Bit, OpSize::i32Bit, 3, 3, ZeroVec, Dest);
+    auto Xar = _VXar(OpSize::i128Bit, OpSize::i32Bit, ZeroVec, Tmp, 2);
+    Result = _VAdd(OpSize::i128Bit, OpSize::i32Bit, Src, Xar);
+  } else {
+    // ARMv8 SHA1 extension provides a `SHA1H` instruction which does a fixed rotate by 30.
+    // This only operates on element 0 rather than element 3. We don't have the luxury of rewriting the x86 SHA algorithm to take advantage of this.
+    // Move the element to zero, rotate, and then move back (Using duplicates).
+    // Saves one instruction versus that path that doesn't support SHA extension.
+    auto Duplicated = _VDupElement(OpSize::i128Bit, OpSize::i32Bit, Dest, 3);
+    auto Sha1HRotated = _VSha1H(Duplicated);
+    auto RotatedNode = _VDupElement(OpSize::i128Bit, OpSize::i32Bit, Sha1HRotated, 0);
+    auto Tmp = _VAdd(OpSize::i128Bit, OpSize::i32Bit, Src, RotatedNode);
+    Result = _VInsElement(OpSize::i128Bit, OpSize::i32Bit, 3, 3, Src, Tmp);
+  }
 
   StoreResult_WithAVXInsert(VectorOpType::SSE, RegClass::FPR, Op, Result);
 }

FEXCore/Source/Interface/IR/IR.json

@@ -2184,6 +2184,18 @@
         ]
       },
 
+      "FPR = VXar OpSize:#RegisterSize, OpSize:#ElementSize, FPR:$LHS, FPR:$RHS, u8:$Rotate": {
+        "Desc": [
+          "Performs an XOR of corresponding elements and then rotates them right by",
+          "an amount between [1, ElementSize]"
+        ],
+        "DestSize": "RegisterSize",
+        "ElementSize": "ElementSize",
+        "EmitValidation": [
+          "RegisterSize == IR::OpSize::i256Bit || RegisterSize == IR::OpSize::i128Bit"
+        ]
+      },
+
       "FPR = VUQAdd OpSize:#RegisterSize, OpSize:#ElementSize, FPR:$Vector1, FPR:$Vector2": {
         "DestSize": "RegisterSize",
         "ElementSize": "ElementSize"

unittests/InstructionCountCI/Crypto/H0F38_SVE128.json

@@ -0,0 +1,28 @@
+{
+  "Features": {
+    "Bitness": 64,
+    "EnabledHostFeatures": [
+      "CRYPTO",
+      "SVE128"
+    ],
+    "DisabledHostFeatures": [
+      "SVE256",
+      "AFP"
+    ]
+  },
+  "Instructions": {
+    "sha1nexte xmm0, xmm1": {
+      "ExpectedInstructionCount": 5,
+      "Comment": [
+        "0x66 0x0f 0x38 0xc8"
+      ],
+      "ExpectedArm64ASM": [
+        "movi v2.2d, #0x0",
+        "mov v3.16b, v2.16b",
+        "mov v3.s[3], v16.s[3]",
+        "xar z2.s, z2.s, z3.s, #2",
+        "add v16.4s, v17.4s, v2.4s"
+      ]
+    }
+  }
+}