From 54451ac731149641969dfc54b40745e312b59c2a Mon Sep 17 00:00:00 2001
From: JosJuice <josjuice@gmail.com>
Date: Sun, 21 Feb 2021 11:21:02 +0100
Subject: [PATCH] JitArm64: Use ConvertSingleToDoubleLower in RW when faster

---
 Source/Core/Core/PowerPC/JitArm64/Jit.h       |  2 ++
 .../JitArm64/JitArm64_FloatingPoint.cpp       |  5 +++++
 .../PowerPC/JitArm64/JitArm64_RegCache.cpp    | 22 +++++++++++++++++--
 3 files changed, 27 insertions(+), 2 deletions(-)

diff --git a/Source/Core/Core/PowerPC/JitArm64/Jit.h b/Source/Core/Core/PowerPC/JitArm64/Jit.h
index 815e808847..2ec565943b 100644
--- a/Source/Core/Core/PowerPC/JitArm64/Jit.h
+++ b/Source/Core/Core/PowerPC/JitArm64/Jit.h
@@ -163,6 +163,8 @@ public:
                                  Arm64Gen::ARM64Reg src_reg,
                                  Arm64Gen::ARM64Reg scratch_reg = Arm64Gen::ARM64Reg::INVALID_REG);
 
+  bool IsFPRStoreSafe(size_t guest_reg) const;
+
 protected:
   struct SlowmemHandler
   {
diff --git a/Source/Core/Core/PowerPC/JitArm64/JitArm64_FloatingPoint.cpp b/Source/Core/Core/PowerPC/JitArm64/JitArm64_FloatingPoint.cpp
index 8d3553afd7..7ba7dd12ea 100644
--- a/Source/Core/Core/PowerPC/JitArm64/JitArm64_FloatingPoint.cpp
+++ b/Source/Core/Core/PowerPC/JitArm64/JitArm64_FloatingPoint.cpp
@@ -572,3 +572,8 @@ void JitArm64::ConvertSingleToDoublePair(size_t guest_reg, ARM64Reg dest_reg, AR
     SetJumpTarget(continue1);
   }
 }
+
+bool JitArm64::IsFPRStoreSafe(size_t guest_reg) const
+{
+  return js.fpr_is_store_safe[guest_reg];
+}
diff --git a/Source/Core/Core/PowerPC/JitArm64/JitArm64_RegCache.cpp b/Source/Core/Core/PowerPC/JitArm64/JitArm64_RegCache.cpp
index 3715c897d2..fdcb55cfb9 100644
--- a/Source/Core/Core/PowerPC/JitArm64/JitArm64_RegCache.cpp
+++ b/Source/Core/Core/PowerPC/JitArm64/JitArm64_RegCache.cpp
@@ -602,9 +602,27 @@ ARM64Reg Arm64FPRCache::RW(size_t preg, RegType type)
     switch (reg.GetType())
     {
     case RegType::Single:
+      // For a store-safe register, conversion is just one instruction regardless of whether
+      // we're whether we're converting a pair, so ConvertSingleToDoublePair followed by a
+      // 128-bit store is faster than INS followed by ConvertSingleToDoubleLower and a
+      // 64-bit store. But for registers which are not store-safe, the latter is better.
       flush_reg = GetReg();
-      m_jit->ConvertSingleToDoublePair(preg, flush_reg, host_reg, flush_reg);
-      [[fallthrough]];
+      if (!m_jit->IsFPRStoreSafe(preg))
+      {
+        ARM64Reg scratch_reg = GetReg();
+        m_float_emit->INS(32, flush_reg, 0, host_reg, 1);
+        m_jit->ConvertSingleToDoubleLower(preg, flush_reg, flush_reg, scratch_reg);
+        m_float_emit->STR(64, IndexType::Unsigned, flush_reg, PPC_REG, u32(PPCSTATE_OFF_PS1(preg)));
+        Unlock(scratch_reg);
+        break;
+      }
+      else
+      {
+        m_jit->ConvertSingleToDoublePair(preg, flush_reg, host_reg, flush_reg);
+        m_float_emit->STR(128, IndexType::Unsigned, flush_reg, PPC_REG,
+                          u32(PPCSTATE_OFF_PS0(preg)));
+      }
+      break;
     case RegType::Register:
       // We are doing a full 128bit store because it takes 2 cycles on a Cortex-A57 to do a 128bit
       // store.