JitArm64: Optimize ConvertDoubleToSingle

3 files changed, 43 insertions, 10 deletions

diff --git a/Source/Core/Core/PowerPC/JitArm64/Jit.h b/Source/Core/Core/PowerPC/JitArm64/Jit.h
index 4f8ea466f4..9d98910660 100644
--- a/Source/Core/Core/PowerPC/JitArm64/Jit.h
+++ b/Source/Core/Core/PowerPC/JitArm64/Jit.h
@@ -220,6 +220,8 @@ private:
   // AsmRoutines
   void GenerateAsm();
   void GenerateCommonAsm();
+  void GenerateConvertDoubleToSingle();
+  void GenerateQuantizedLoadStores();
 
   // Profiling
   void BeginTimeProfile(JitBlock* b);
diff --git a/Source/Core/Core/PowerPC/JitArm64/JitArm64_FloatingPoint.cpp b/Source/Core/Core/PowerPC/JitArm64/JitArm64_FloatingPoint.cpp
index 4a4830ef63..50c9d1f85c 100644
--- a/Source/Core/Core/PowerPC/JitArm64/JitArm64_FloatingPoint.cpp
+++ b/Source/Core/Core/PowerPC/JitArm64/JitArm64_FloatingPoint.cpp
@@ -397,12 +397,12 @@ void JitArm64::ConvertDoubleToSingleLower(ARM64Reg dest_reg, ARM64Reg src_reg)
 {
   FlushCarry();
 
-  const BitSet32 gpr_saved = gpr.GetCallerSavedUsed();
+  const BitSet32 gpr_saved = gpr.GetCallerSavedUsed() & BitSet32{0, 1, 2, 3, 30};
   ABI_PushRegisters(gpr_saved);
 
   m_float_emit.UMOV(64, ARM64Reg::X0, src_reg, 0);
-  QuickCallFunction(ARM64Reg::X1, &ConvertToSingle);
-  m_float_emit.INS(32, dest_reg, 0, ARM64Reg::W0);
+  BL(cdts);
+  m_float_emit.INS(32, dest_reg, 0, ARM64Reg::W1);
 
   ABI_PopRegisters(gpr_saved);
 }
@@ -411,16 +411,16 @@ void JitArm64::ConvertDoubleToSinglePair(ARM64Reg dest_reg, ARM64Reg src_reg)
 {
   FlushCarry();
 
-  const BitSet32 gpr_saved = gpr.GetCallerSavedUsed();
+  const BitSet32 gpr_saved = gpr.GetCallerSavedUsed() & BitSet32{0, 1, 2, 3, 30};
   ABI_PushRegisters(gpr_saved);
 
   m_float_emit.UMOV(64, ARM64Reg::X0, src_reg, 0);
-  QuickCallFunction(ARM64Reg::X1, &ConvertToSingle);
-  m_float_emit.INS(32, dest_reg, 0, ARM64Reg::W0);
+  BL(cdts);
+  m_float_emit.INS(32, dest_reg, 0, ARM64Reg::W1);
 
   m_float_emit.UMOV(64, ARM64Reg::X0, src_reg, 1);
-  QuickCallFunction(ARM64Reg::X1, &ConvertToSingle);
-  m_float_emit.INS(32, dest_reg, 1, ARM64Reg::W0);
+  BL(cdts);
+  m_float_emit.INS(32, dest_reg, 1, ARM64Reg::W1);
 
   ABI_PopRegisters(gpr_saved);
 }
diff --git a/Source/Core/Core/PowerPC/JitArm64/JitAsm.cpp b/Source/Core/Core/PowerPC/JitArm64/JitAsm.cpp
index 3b426a6214..c686c31ce4 100644
--- a/Source/Core/Core/PowerPC/JitArm64/JitAsm.cpp
+++ b/Source/Core/Core/PowerPC/JitArm64/JitAsm.cpp
@@ -195,6 +195,39 @@ void JitArm64::GenerateAsm()
 
 void JitArm64::GenerateCommonAsm()
 {
+  GetAsmRoutines()->cdts = GetCodePtr();
+  GenerateConvertDoubleToSingle();
+  JitRegister::Register(GetAsmRoutines()->cdts, GetCodePtr(), "JIT_cdts");
+
+  GenerateQuantizedLoadStores();
+}
+
+// Input in X0, output in W1, clobbers X0-X3 and flags.
+void JitArm64::GenerateConvertDoubleToSingle()
+{
+  UBFX(ARM64Reg::X2, ARM64Reg::X0, 52, 11);
+  SUB(ARM64Reg::W3, ARM64Reg::W2, 874);
+  CMP(ARM64Reg::W3, 896 - 874);
+  LSR(ARM64Reg::X1, ARM64Reg::X0, 32);
+  FixupBranch denormal = B(CCFlags::CC_LS);
+
+  ANDI2R(ARM64Reg::X1, ARM64Reg::X1, 0xc0000000);
+  BFXIL(ARM64Reg::X1, ARM64Reg::X0, 29, 30);
+  RET();
+
+  SetJumpTarget(denormal);
+  LSR(ARM64Reg::X3, ARM64Reg::X0, 21);
+  MOVZ(ARM64Reg::X0, 905);
+  ORRI2R(ARM64Reg::W3, ARM64Reg::W3, 0x80000000);
+  SUB(ARM64Reg::W2, ARM64Reg::W0, ARM64Reg::W2);
+  LSRV(ARM64Reg::W2, ARM64Reg::W3, ARM64Reg::W2);
+  ANDI2R(ARM64Reg::X3, ARM64Reg::X1, 0x80000000);
+  ORR(ARM64Reg::X1, ARM64Reg::X3, ARM64Reg::X2);
+  RET();
+}
+
+void JitArm64::GenerateQuantizedLoadStores()
+{
   // X0 is the scale
   // X1 is address
   // X2 is a temporary on stores
@@ -654,6 +687,4 @@ void JitArm64::GenerateCommonAsm()
   paired_store_quantized[29] = storeSingleU16Slow;
   paired_store_quantized[30] = storeSingleS8Slow;
   paired_store_quantized[31] = storeSingleS16Slow;
-
-  GetAsmRoutines()->mfcr = nullptr;
 }