From b79f33526dbe0fe293f3940d1eeb6d916905fee6 Mon Sep 17 00:00:00 2001
From: chrisps <chss95cs@gmail.com>
Date: Wed, 15 Jan 2020 15:57:09 -0800
Subject: [PATCH] Optimized CONVERT_I64_TO_F64 with neat overflow trick

Reduced instruction count from 11 to 8, eliminated a movq stall.
---
 src/xenia/cpu/backend/x64/x64_sequences.cc | 31 ++++++++--------------
 1 file changed, 11 insertions(+), 20 deletions(-)

diff --git a/src/xenia/cpu/backend/x64/x64_sequences.cc b/src/xenia/cpu/backend/x64/x64_sequences.cc
index 70956b52b..d91a674f1 100644
--- a/src/xenia/cpu/backend/x64/x64_sequences.cc
+++ b/src/xenia/cpu/backend/x64/x64_sequences.cc
@@ -317,31 +317,22 @@ struct CONVERT_I32_F64
 struct CONVERT_I64_F64
     : Sequence<CONVERT_I64_F64, I<OPCODE_CONVERT, I64Op, F64Op>> {
   static void Emit(X64Emitter& e, const EmitArgType& i) {
-    // Copy src1.
-    e.movq(e.rcx, i.src1);
 
-    // TODO(benvanik): saturation check? cvtt* (trunc?)
+    e.xor_ (e.eax, e.eax);
+    
+    e.vcomisd(i.src1, e.GetXmmConstPtr(XmmConst::XMMZero));
     if (i.instr->flags == ROUND_TO_ZERO) {
       e.vcvttsd2si(i.dest, i.src1);
-    } else {
+    }
+    else {
       e.vcvtsd2si(i.dest, i.src1);
     }
-
-    // 0x8000000000000000
-    e.mov(e.rax, 0x1);
-    e.shl(e.rax, 63);
-
-    // Saturate positive overflow
-    // TODO(DrChat): Find a shorter equivalent sequence.
-    // if (result ind. && src1 >= 0)
-    //   result = 0x7FFFFFFFFFFFFFFF;
-    e.cmp(e.rax, i.dest);
-    e.sete(e.al);
-    e.movzx(e.rax, e.al);
-    e.shr(e.rcx, 63);
-    e.xor_(e.rcx, 0x01);
-    e.and_(e.rax, e.rcx);
-
+    // cf set if less than
+    e.setnc(e.cl);
+    e.cmp(i.dest, -1LL);
+    // if dest == 0x80000000 and not inp < 0 then dest = 0x7FFFFFFF
+    e.seto(e.al);
+    e.and_ (e.al, e.cl);
     e.sub(i.dest, e.rax);
   }
 };