From b90757651076cfb5122af387fa73904a0b86f231 Mon Sep 17 00:00:00 2001
From: Ryan Houdek <Sonicadvance1@gmail.com>
Date: Fri, 7 Aug 2015 02:32:16 -0500
Subject: [PATCH] [AArch64] Support profiling by cycle counters if they are
 available to EL0

---
 Source/Core/Common/Arm64Emitter.cpp       |  6 +++
 Source/Core/Common/Arm64Emitter.h         |  3 ++
 Source/Core/Core/PowerPC/JitArm64/Jit.cpp | 66 ++++++++++++++++++++---
 Source/Core/Core/PowerPC/JitArm64/Jit.h   |  6 +++
 4 files changed, 73 insertions(+), 8 deletions(-)

diff --git a/Source/Core/Common/Arm64Emitter.cpp b/Source/Core/Common/Arm64Emitter.cpp
index 9069bedeb7..3a8303c2e3 100644
--- a/Source/Core/Common/Arm64Emitter.cpp
+++ b/Source/Core/Common/Arm64Emitter.cpp
@@ -1102,6 +1102,12 @@ static void GetSystemReg(PStateField field, int &o0, int &op1, int &CRn, int &CR
 	case FIELD_FPSR:
 		o0 = 3; op1 = 3; CRn = 4; CRm = 4; op2 = 1;
 		break;
+	case FIELD_PMCR_EL0:
+		o0 = 3; op1 = 3; CRn = 9; CRm = 6; op2 = 0;
+		break;
+	case FIELD_PMCCNTR_EL0:
+		o0 = 3; op1 = 3; CRn = 9; CRm = 7; op2 = 0;
+		break;
 	default:
 		_assert_msg_(DYNA_REC, false, "Invalid PStateField to do a register move from/to");
 		break;
diff --git a/Source/Core/Common/Arm64Emitter.h b/Source/Core/Common/Arm64Emitter.h
index 0133d7f0d4..f6cb834a5d 100644
--- a/Source/Core/Common/Arm64Emitter.h
+++ b/Source/Core/Common/Arm64Emitter.h
@@ -172,6 +172,8 @@ enum PStateField
 	FIELD_DAIFSet,
 	FIELD_DAIFClr,
 	FIELD_NZCV,	// The only system registers accessible from EL0 (user space)
+	FIELD_PMCR_EL0,
+	FIELD_PMCCNTR_EL0,
 	FIELD_FPCR = 0x340,
 	FIELD_FPSR = 0x341,
 };
@@ -809,6 +811,7 @@ public:
 	void FCVTL(u8 size, ARM64Reg Rd, ARM64Reg Rn);
 	void FCVTL2(u8 size, ARM64Reg Rd, ARM64Reg Rn);
 	void FCVTN(u8 dest_size, ARM64Reg Rd, ARM64Reg Rn);
+	void FCVTN2(u8 dest_size, ARM64Reg Rd, ARM64Reg Rn);
 	void FCVTZS(u8 size, ARM64Reg Rd, ARM64Reg Rn);
 	void FCVTZU(u8 size, ARM64Reg Rd, ARM64Reg Rn);
 	void FDIV(u8 size, ARM64Reg Rd, ARM64Reg Rn, ARM64Reg Rm);
diff --git a/Source/Core/Core/PowerPC/JitArm64/Jit.cpp b/Source/Core/Core/PowerPC/JitArm64/Jit.cpp
index 2c002f9dc7..64ae021930 100644
--- a/Source/Core/Core/PowerPC/JitArm64/Jit.cpp
+++ b/Source/Core/Core/PowerPC/JitArm64/Jit.cpp
@@ -15,6 +15,15 @@
 using namespace Arm64Gen;
 
 static const int AARCH64_FARCODE_SIZE = 1024 * 1024 * 16;
+static bool HasCycleCounters()
+{
+	// Bit needs to be set to support cycle counters
+	const u32 PMUSERENR_CR = 0x4;
+	u32 reg;
+	asm ("mrs %[val], PMUSERENR_EL0"
+			: [val] "=r" (reg));
+	return !!(reg & PMUSERENR_CR);
+}
 
 void JitArm64::Init()
 {
@@ -34,6 +43,8 @@ void JitArm64::Init()
 	code_block.m_gpa = &js.gpa;
 	code_block.m_fpa = &js.fpa;
 	analyzer.SetOption(PPCAnalyst::PPCAnalyzer::OPTION_CONDITIONAL_CONTINUE);
+
+	m_supports_cycle_counter = HasCycleCounters();
 }
 
 void JitArm64::ClearCache()
@@ -233,26 +244,65 @@ void JitArm64::DumpCode(const u8* start, const u8* end)
 	WARN_LOG(DYNA_REC, "Code dump from %p to %p:\n%s", start, end, output.c_str());
 }
 
+void JitArm64::EmitResetCycleCounters()
+{
+	const u32 PMCR_EL0_E  = 1;
+	const u32 PMCR_EL0_P  = 2;
+	const u32 PMCR_EL0_C  = 4;
+	const u32 PMCR_EL0_LC = 0x40;
+	_MSR(FIELD_PMCR_EL0, X0);
+	MOVI2R(X1, PMCR_EL0_E |
+	           PMCR_EL0_P |
+	           PMCR_EL0_C |
+	           PMCR_EL0_LC);
+	ORR(X0, X0, X1);
+	MRS(X0, FIELD_PMCR_EL0);
+}
+
+void JitArm64::EmitGetCycles(Arm64Gen::ARM64Reg reg)
+{
+	_MSR(FIELD_PMCCNTR_EL0, reg);
+}
+
 void JitArm64::BeginTimeProfile(JitBlock* b)
 {
 	b->ticCounter = 0;
 	b->ticStart = 0;
 	b->ticStop = 0;
 
-	MOVI2R(X1, (u64)QueryPerformanceCounter);
-	MOVI2R(X0, (u64)&b->ticStart);
-	BLR(X1);
+	if (m_supports_cycle_counter)
+	{
+		EmitResetCycleCounters();
+		EmitGetCycles(X1);
+		MOVI2R(X0, (u64)&b->ticStart);
+		STR(INDEX_UNSIGNED, X1, X0, 0);
+	}
+	else
+	{
+		MOVI2R(X1, (u64)QueryPerformanceCounter);
+		MOVI2R(X0, (u64)&b->ticStart);
+		BLR(X1);
+	}
 }
 
 void JitArm64::EndTimeProfile(JitBlock* b)
 {
-	MOVI2R(X1, (u64)QueryPerformanceCounter);
-	MOVI2R(X0, (u64)&b->ticStop);
-	BLR(X1);
+	if (m_supports_cycle_counter)
+	{
+		EmitGetCycles(X2);
+		MOVI2R(X0, (u64)&b->ticStart);
+	}
+	else
+	{
+		MOVI2R(X1, (u64)QueryPerformanceCounter);
+		MOVI2R(X0, (u64)&b->ticStop);
+		BLR(X1);
+
+		MOVI2R(X0, (u64)&b->ticStart);
+		LDR(INDEX_UNSIGNED, X2, X0, 8); // Stop
+	}
 
-	MOVI2R(X0, (u64)&b->ticStart);
 	LDR(INDEX_UNSIGNED, X1, X0, 0); // Start
-	LDR(INDEX_UNSIGNED, X2, X0, 8); // Stop
 	LDR(INDEX_UNSIGNED, X3, X0, 16); // Counter
 	SUB(X2, X2, X1);
 	ADD(X3, X3, X2);
diff --git a/Source/Core/Core/PowerPC/JitArm64/Jit.h b/Source/Core/Core/PowerPC/JitArm64/Jit.h
index 1d7276bc0b..881626491b 100644
--- a/Source/Core/Core/PowerPC/JitArm64/Jit.h
+++ b/Source/Core/Core/PowerPC/JitArm64/Jit.h
@@ -238,6 +238,12 @@ private:
 	Arm64Gen::ARM64CodeBlock farcode;
 	u8* nearcode; // Backed up when we switch to far code.
 
+	// Do we support cycle counter profiling?
+	bool m_supports_cycle_counter;
+
+	void EmitResetCycleCounters();
+	void EmitGetCycles(Arm64Gen::ARM64Reg reg);
+
 	// Simple functions to switch between near and far code emitting
 	void SwitchToFarCode()
 	{