From f4ce6fb621a97cdbe0cbce4bf9224d317f5fcfb9 Mon Sep 17 00:00:00 2001
From: Alyssa Rosenzweig <alyssa@rosenzweig.io>
Date: Tue, 26 Nov 2024 08:38:20 -0500
Subject: [PATCH] OpcodeDispatcher: optimize AAS/AAD flag

Signed-off-by: Alyssa Rosenzweig <alyssa@rosenzweig.io>
---
 .../Source/Interface/Core/OpcodeDispatcher.cpp  |  6 ++----
 .../Source/Interface/Core/OpcodeDispatcher.h    | 17 +++++++++++++++++
 2 files changed, 19 insertions(+), 4 deletions(-)

diff --git a/FEXCore/Source/Interface/Core/OpcodeDispatcher.cpp b/FEXCore/Source/Interface/Core/OpcodeDispatcher.cpp
index fab35fb0b..f6d79666a 100644
--- a/FEXCore/Source/Interface/Core/OpcodeDispatcher.cpp
+++ b/FEXCore/Source/Interface/Core/OpcodeDispatcher.cpp
@@ -2861,8 +2861,7 @@ void OpDispatchBuilder::AAAOp(OpcodeArgs) {
   auto AF = CalculateAFForDecimal(A);
 
   // CF = AF, OF/SF/ZF/PF undefined
-  ZeroNZCV();
-  SetCFDirect(AF);
+  SetCFDirect_InvalidateNZV(AF);
   SetAFAndFixup(AF);
   CalculateDeferredFlags();
 
@@ -2881,8 +2880,7 @@ void OpDispatchBuilder::AASOp(OpcodeArgs) {
   auto AF = CalculateAFForDecimal(A);
 
   // CF = AF, OF/SF/ZF/PF undefined
-  ZeroNZCV();
-  SetCFDirect(AF);
+  SetCFDirect_InvalidateNZV(AF);
   SetAFAndFixup(AF);
   CalculateDeferredFlags();
 
diff --git a/FEXCore/Source/Interface/Core/OpcodeDispatcher.h b/FEXCore/Source/Interface/Core/OpcodeDispatcher.h
index 118cd10bf..6cd14e223 100644
--- a/FEXCore/Source/Interface/Core/OpcodeDispatcher.h
+++ b/FEXCore/Source/Interface/Core/OpcodeDispatcher.h
@@ -1751,6 +1751,23 @@ private:
     CFInverted = true;
   }
 
+  // Set CF directly to the given 0/1 value. This needs to respect the
+  // invert. We use a subtraction:
+  //
+  //     0 - x = 0 + (~x) + 1.
+  //
+  // If x = 0, then 0 + (~0) + 1 = 0x100000000 so hardware C is set.
+  // If x = 1, then 0 + (~1) + 1 = 0x0ffffffff so hardware C is not set.
+  void SetCFDirect_InvalidateNZV(Ref Value, unsigned ValueOffset = 0, bool MustMask = false) {
+    if (ValueOffset || MustMask) {
+      Value = _Bfe(OpSize::i64Bit, 1, ValueOffset, Value);
+    }
+
+    HandleNZCVWrite();
+    _SubNZCV(OpSize::i32Bit, _Constant(0), Value);
+    CFInverted = true;
+  }
+
   void SetCFInverted(Ref Value, unsigned ValueOffset = 0, bool MustMask = false) {
     SetRFLAG(Value, X86State::RFLAG_CF_RAW_LOC, ValueOffset, MustMask);
     CFInverted = true;