From f695a215ad84607a2df8f31f2138918926eb3f0c Mon Sep 17 00:00:00 2001
From: riperiperi <rhy3756547@hotmail.com>
Date: Wed, 25 Mar 2020 06:20:29 +0000
Subject: Add Fast Paths for Crypto instructions (A32/A64) (#1026)

* Add Fast Paths for Crypto instructions (A32/A64)

* Replace additional XOR with passing in const zero.
---
 ARMeilleure/Instructions/InstEmitSimdCrypto32.cs | 54 ++++++++++++++++++++++--
 1 file changed, 50 insertions(+), 4 deletions(-)

(limited to 'ARMeilleure/Instructions/InstEmitSimdCrypto32.cs')

diff --git a/ARMeilleure/Instructions/InstEmitSimdCrypto32.cs b/ARMeilleure/Instructions/InstEmitSimdCrypto32.cs
index 1cfce3ad..f62fd307 100644
--- a/ARMeilleure/Instructions/InstEmitSimdCrypto32.cs
+++ b/ARMeilleure/Instructions/InstEmitSimdCrypto32.cs
@@ -15,7 +15,17 @@ namespace ARMeilleure.Instructions
             Operand d = GetVecA32(op.Qd);
             Operand n = GetVecA32(op.Qm);
 
-            context.Copy(d, context.Call(new _V128_V128_V128(SoftFallback.Decrypt), d, n));
+            Operand res;
+            if (Optimizations.UseAesni)
+            {
+                res = context.AddIntrinsic(Intrinsic.X86Aesdeclast, context.AddIntrinsic(Intrinsic.X86Xorpd, d, n), context.VectorZero());
+            }
+            else
+            {
+                res = context.Call(new _V128_V128_V128(SoftFallback.Decrypt), d, n);
+            }
+
+            context.Copy(d, res);
         }
 
         public static void Aese_V(ArmEmitterContext context)
@@ -25,7 +35,17 @@ namespace ARMeilleure.Instructions
             Operand d = GetVecA32(op.Qd);
             Operand n = GetVecA32(op.Qm);
 
-            context.Copy(d, context.Call(new _V128_V128_V128(SoftFallback.Encrypt), d, n));
+            Operand res;
+            if (Optimizations.UseAesni)
+            {
+                res = context.AddIntrinsic(Intrinsic.X86Aesenclast, context.AddIntrinsic(Intrinsic.X86Xorpd, d, n), context.VectorZero());
+            }
+            else
+            {
+                res = context.Call(new _V128_V128_V128(SoftFallback.Encrypt), d, n);
+            }
+
+            context.Copy(d, res);
         }
 
         public static void Aesimc_V(ArmEmitterContext context)
@@ -34,7 +54,17 @@ namespace ARMeilleure.Instructions
 
             Operand n = GetVecA32(op.Qm);
 
-            context.Copy(GetVec(op.Qd), context.Call(new _V128_V128(SoftFallback.InverseMixColumns), n));
+            Operand res;
+            if (Optimizations.UseAesni)
+            {
+                res = context.AddIntrinsic(Intrinsic.X86Aesimc, n);
+            }
+            else
+            {
+                res = context.Call(new _V128_V128(SoftFallback.InverseMixColumns), n);
+            }
+
+            context.Copy(GetVecA32(op.Qd), res);
         }
 
         public static void Aesmc_V(ArmEmitterContext context)
@@ -43,7 +73,23 @@ namespace ARMeilleure.Instructions
 
             Operand n = GetVecA32(op.Qm);
 
-            context.Copy(GetVec(op.Qd), context.Call(new _V128_V128(SoftFallback.MixColumns), n));
+            Operand res;
+            if (Optimizations.UseAesni)
+            {
+                Operand roundKey = context.VectorZero();
+
+                // Inverse Shift Rows, Inverse Sub Bytes, xor 0 so nothing happens.
+                res = context.AddIntrinsic(Intrinsic.X86Aesdeclast, n, roundKey);
+
+                // Shift Rows, Sub Bytes, Mix Columns (!), xor 0 so nothing happens.
+                res = context.AddIntrinsic(Intrinsic.X86Aesenc, res, roundKey);
+            }
+            else
+            {
+                res = context.Call(new _V128_V128(SoftFallback.MixColumns), n);
+            }
+
+            context.Copy(GetVecA32(op.Qd), res);
         }
     }
 }
-- 
cgit v1.2.3