From 5e806eb7b27db99510fd565505587b7d7c246cb6 Mon Sep 17 00:00:00 2001
From: nodchip <nodchip@gmail.com>
Date: Thu, 23 Sep 2010 02:05:04 +0000
Subject: [PATCH] Jit64/JitIL: Implemented simple function inlining for
 speedup.

git-svn-id: https://dolphin-emu.googlecode.com/svn/trunk@6223 8ced0084-cf51-0410-be5f-012b33b47a6e
---
 .../Core/Src/PowerPC/Jit64/Jit_Branch.cpp     |  7 ++++++
 .../Core/Src/PowerPC/Jit64IL/JitIL_Branch.cpp |  8 ++++++
 Source/Core/Core/Src/PowerPC/PPCAnalyst.cpp   | 25 +++++++++++++++++++
 3 files changed, 40 insertions(+)

diff --git a/Source/Core/Core/Src/PowerPC/Jit64/Jit_Branch.cpp b/Source/Core/Core/Src/PowerPC/Jit64/Jit_Branch.cpp
index aef25e9779..34f9839d6f 100644
--- a/Source/Core/Core/Src/PowerPC/Jit64/Jit_Branch.cpp
+++ b/Source/Core/Core/Src/PowerPC/Jit64/Jit_Branch.cpp
@@ -221,6 +221,13 @@ void Jit64::bclrx(UGeckoInstruction inst)
 	INSTRUCTION_START
 	JITDISABLE(Branch)
 
+	if (!js.isLastInstruction &&
+		(inst.BO & (1 << 4)) && (inst.BO & (1 << 2))) {
+		if (inst.LK)
+			MOV(32, M(&LR), Imm32(js.compilerPC + 4));
+		return;
+	}
+
 	gpr.Flush(FLUSH_ALL);
 	fpr.Flush(FLUSH_ALL);
 
diff --git a/Source/Core/Core/Src/PowerPC/Jit64IL/JitIL_Branch.cpp b/Source/Core/Core/Src/PowerPC/Jit64IL/JitIL_Branch.cpp
index 608d362032..207ac60a5f 100644
--- a/Source/Core/Core/Src/PowerPC/Jit64IL/JitIL_Branch.cpp
+++ b/Source/Core/Core/Src/PowerPC/Jit64IL/JitIL_Branch.cpp
@@ -185,6 +185,14 @@ void JitIL::bcctrx(UGeckoInstruction inst)
 void JitIL::bclrx(UGeckoInstruction inst)
 {
 	NORMALBRANCH_START
+
+	if (!js.isLastInstruction &&
+		(inst.BO & (1 << 4)) && (inst.BO & (1 << 2))) {
+		if (inst.LK)
+			ibuild.EmitStoreLink(ibuild.EmitIntConst(js.compilerPC + 4));
+		return;
+	}
+
 	if (inst.hex == 0x4e800020) {
 		ibuild.EmitBranchUncond(ibuild.EmitLoadLink());
 		return;
diff --git a/Source/Core/Core/Src/PowerPC/PPCAnalyst.cpp b/Source/Core/Core/Src/PowerPC/PPCAnalyst.cpp
index 5de88c305c..bb94789588 100644
--- a/Source/Core/Core/Src/PowerPC/PPCAnalyst.cpp
+++ b/Source/Core/Core/Src/PowerPC/PPCAnalyst.cpp
@@ -325,6 +325,8 @@ u32 Flatten(u32 address, int *realsize, BlockStats *st, BlockRegStats *gpa,
 	CodeOp *code = buffer->codebuffer;
 	bool foundExit = false;
 
+	u32 returnAddress = 0;
+
 	// Do analysis of the code, look for dependencies etc
 	int numSystemInstructions = 0;
 	for (int i = 0; i < maxsize; i++)
@@ -454,6 +456,29 @@ u32 Flatten(u32 address, int *realsize, BlockStats *st, BlockRegStats *gpa,
 				if (destination != blockstart)
 					follow = true;
 			}
+			else if (inst.OPCD == 19 && inst.SUBOP10 == 16 &&
+				(inst.BO & (1 << 4)) && (inst.BO & (1 << 2)) &&
+				returnAddress != 0)
+			{
+				// bclrx with unconditional branch = return
+				follow = true;
+				destination = returnAddress;
+				returnAddress = 0;
+
+				if (inst.LK)
+					returnAddress = address + 4;
+			}
+			else if (inst.OPCD == 31 && inst.SUBOP10 == 467)
+			{
+				// mtspr
+				const u32 index = (inst.SPRU << 5) | (inst.SPRL & 0x1F);
+				if (index == SPR_LR) {
+					// We give up to follow the return address
+					// because we have to check the register usage.
+					returnAddress = 0;
+				}
+			}
+
 			if (follow)
 				numFollows++;
 			// TODO: Find the optimal value for FUNCTION_FOLLOWING_THRESHOLD.