From 018e24762481fd59e9c3e9e0bf2a52dddfdecdca Mon Sep 17 00:00:00 2001
From: JosJuice <josjuice@gmail.com>
Date: Sun, 31 Jan 2021 19:20:02 +0100
Subject: [PATCH] JitArm64: Optimize ConvertSingleToDouble, part 1

---
 Source/Core/Core/PowerPC/JitArm64/Jit.h       |  1 +
 .../JitArm64/JitArm64_FloatingPoint.cpp       | 15 +++---
 Source/Core/Core/PowerPC/JitArm64/JitAsm.cpp  | 46 +++++++++++++++++++
 .../Core/PowerPC/JitCommon/JitAsmCommon.h     |  1 +
 4 files changed, 54 insertions(+), 9 deletions(-)

diff --git a/Source/Core/Core/PowerPC/JitArm64/Jit.h b/Source/Core/Core/PowerPC/JitArm64/Jit.h
index 9d98910660..cc65155ccd 100644
--- a/Source/Core/Core/PowerPC/JitArm64/Jit.h
+++ b/Source/Core/Core/PowerPC/JitArm64/Jit.h
@@ -221,6 +221,7 @@ private:
   void GenerateAsm();
   void GenerateCommonAsm();
   void GenerateConvertDoubleToSingle();
+  void GenerateConvertSingleToDouble();
   void GenerateQuantizedLoadStores();
 
   // Profiling
diff --git a/Source/Core/Core/PowerPC/JitArm64/JitArm64_FloatingPoint.cpp b/Source/Core/Core/PowerPC/JitArm64/JitArm64_FloatingPoint.cpp
index 50c9d1f85c..d3d5f7ddbf 100644
--- a/Source/Core/Core/PowerPC/JitArm64/JitArm64_FloatingPoint.cpp
+++ b/Source/Core/Core/PowerPC/JitArm64/JitArm64_FloatingPoint.cpp
@@ -9,7 +9,6 @@
 #include "Core/ConfigManager.h"
 #include "Core/Core.h"
 #include "Core/CoreTiming.h"
-#include "Core/PowerPC/Interpreter/Interpreter_FPUtils.h"
 #include "Core/PowerPC/JitArm64/Jit.h"
 #include "Core/PowerPC/JitArm64/JitArm64_RegCache.h"
 #include "Core/PowerPC/PPCTables.h"
@@ -390,9 +389,6 @@ void JitArm64::fctiwzx(UGeckoInstruction inst)
 // instructions, they must convert floats bitexact and never flush denormals to zero or turn SNaNs
 // into QNaNs. This means we can't just use FCVT/FCVTL/FCVTN.
 
-// When calling the conversion functions, we are cheating a little and not
-// saving the FPRs since we know the functions happen to not use them.
-
 void JitArm64::ConvertDoubleToSingleLower(ARM64Reg dest_reg, ARM64Reg src_reg)
 {
   FlushCarry();
@@ -429,11 +425,11 @@ void JitArm64::ConvertSingleToDoubleLower(ARM64Reg dest_reg, ARM64Reg src_reg)
 {
   FlushCarry();
 
-  const BitSet32 gpr_saved = gpr.GetCallerSavedUsed();
+  const BitSet32 gpr_saved = gpr.GetCallerSavedUsed() & BitSet32{0, 1, 2, 3, 4, 30};
   ABI_PushRegisters(gpr_saved);
 
   m_float_emit.UMOV(32, ARM64Reg::W0, src_reg, 0);
-  QuickCallFunction(ARM64Reg::X1, &ConvertToDouble);
+  BL(cstd);
   m_float_emit.INS(64, dest_reg, 0, ARM64Reg::X0);
 
   ABI_PopRegisters(gpr_saved);
@@ -443,15 +439,16 @@ void JitArm64::ConvertSingleToDoublePair(ARM64Reg dest_reg, ARM64Reg src_reg)
 {
   FlushCarry();
 
-  const BitSet32 gpr_saved = gpr.GetCallerSavedUsed();
+  // Save X0-X4 and X30 if they're in use
+  const BitSet32 gpr_saved = gpr.GetCallerSavedUsed() & BitSet32{0, 1, 2, 3, 4, 30};
   ABI_PushRegisters(gpr_saved);
 
   m_float_emit.UMOV(32, ARM64Reg::W0, src_reg, 1);
-  QuickCallFunction(ARM64Reg::X1, &ConvertToDouble);
+  BL(cstd);
   m_float_emit.INS(64, dest_reg, 1, ARM64Reg::X0);
 
   m_float_emit.UMOV(32, ARM64Reg::W0, src_reg, 0);
-  QuickCallFunction(ARM64Reg::X1, &ConvertToDouble);
+  BL(cstd);
   m_float_emit.INS(64, dest_reg, 0, ARM64Reg::X0);
 
   ABI_PopRegisters(gpr_saved);
diff --git a/Source/Core/Core/PowerPC/JitArm64/JitAsm.cpp b/Source/Core/Core/PowerPC/JitArm64/JitAsm.cpp
index c686c31ce4..57fa59d9df 100644
--- a/Source/Core/Core/PowerPC/JitArm64/JitAsm.cpp
+++ b/Source/Core/Core/PowerPC/JitArm64/JitAsm.cpp
@@ -199,6 +199,10 @@ void JitArm64::GenerateCommonAsm()
   GenerateConvertDoubleToSingle();
   JitRegister::Register(GetAsmRoutines()->cdts, GetCodePtr(), "JIT_cdts");
 
+  GetAsmRoutines()->cstd = GetCodePtr();
+  GenerateConvertSingleToDouble();
+  JitRegister::Register(GetAsmRoutines()->cdts, GetCodePtr(), "JIT_cstd");
+
   GenerateQuantizedLoadStores();
 }
 
@@ -226,6 +230,48 @@ void JitArm64::GenerateConvertDoubleToSingle()
   RET();
 }
 
+// Input in W0, output in X0, clobbers X0-X4 and flags.
+void JitArm64::GenerateConvertSingleToDouble()
+{
+  UBFX(ARM64Reg::W1, ARM64Reg::W0, 23, 8);
+  FixupBranch normal_or_nan = CBNZ(ARM64Reg::W1);
+
+  ANDI2R(ARM64Reg::W1, ARM64Reg::W0, 0x007fffff);
+  FixupBranch denormal = CBNZ(ARM64Reg::W1);
+
+  // Zero
+  LSL(ARM64Reg::X0, ARM64Reg::X0, 32);
+  RET();
+
+  SetJumpTarget(denormal);
+  ANDI2R(ARM64Reg::W2, ARM64Reg::W0, 0x80000000);
+  CLZ(ARM64Reg::X3, ARM64Reg::X1);
+  LSL(ARM64Reg::X2, ARM64Reg::X2, 32);
+  ORRI2R(ARM64Reg::X4, ARM64Reg::X3, 0xffffffffffffffc0);
+  SUB(ARM64Reg::X2, ARM64Reg::X2, ARM64Reg::X3, ArithOption(ARM64Reg::X3, ShiftType::LSL, 52));
+  ADD(ARM64Reg::X3, ARM64Reg::X4, 23);
+  LSLV(ARM64Reg::X1, ARM64Reg::X1, ARM64Reg::X3);
+  BFI(ARM64Reg::X2, ARM64Reg::X1, 30, 22);
+  MOVI2R(ARM64Reg::X1, 0x3a90000000000000);
+  ADD(ARM64Reg::X0, ARM64Reg::X2, ARM64Reg::X1);
+  RET();
+
+  SetJumpTarget(normal_or_nan);
+  CMP(ARM64Reg::W1, 0xff);
+  ANDI2R(ARM64Reg::W2, ARM64Reg::W0, 0x40000000);
+  CSET(ARM64Reg::W4, CCFlags::CC_NEQ);
+  ANDI2R(ARM64Reg::W3, ARM64Reg::W0, 0xc0000000);
+  EOR(ARM64Reg::W2, ARM64Reg::W4, ARM64Reg::W2, ArithOption(ARM64Reg::W2, ShiftType::LSR, 30));
+  MOVI2R(ARM64Reg::X1, 0x3800000000000000);
+  ANDI2R(ARM64Reg::W4, ARM64Reg::W0, 0x3fffffff);
+  LSL(ARM64Reg::X3, ARM64Reg::X3, 32);
+  CMP(ARM64Reg::W2, 0);
+  CSEL(ARM64Reg::X1, ARM64Reg::X1, ARM64Reg::ZR, CCFlags::CC_NEQ);
+  BFI(ARM64Reg::X3, ARM64Reg::X4, 29, 30);
+  ORR(ARM64Reg::X0, ARM64Reg::X3, ARM64Reg::X1);
+  RET();
+}
+
 void JitArm64::GenerateQuantizedLoadStores()
 {
   // X0 is the scale
diff --git a/Source/Core/Core/PowerPC/JitCommon/JitAsmCommon.h b/Source/Core/Core/PowerPC/JitCommon/JitAsmCommon.h
index d8e22a0a3a..c525e7849c 100644
--- a/Source/Core/Core/PowerPC/JitCommon/JitAsmCommon.h
+++ b/Source/Core/Core/PowerPC/JitCommon/JitAsmCommon.h
@@ -26,6 +26,7 @@ struct CommonAsmRoutinesBase
   const u8* fres;
   const u8* mfcr;
   const u8* cdts;
+  const u8* cstd;
 
   // In: array index: GQR to use.
   // In: ECX: Address to read from.