Merge pull request #975 from FioraAeterna/jittimerinline

JIT: completely inline timer reading

3 files changed, 24 insertions, 9 deletions

diff --git a/Source/Core/Core/CoreTiming.cpp b/Source/Core/Core/CoreTiming.cpp
index d64a39b51d..30ac74585a 100644
--- a/Source/Core/Core/CoreTiming.cpp
+++ b/Source/Core/Core/CoreTiming.cpp
@@ -50,13 +50,13 @@ static Event *eventPool = nullptr;
 int slicelength;
 static int maxSliceLength = MAX_SLICE_LENGTH;
 
-static s64 globalTimer;
 static s64 idledCycles;
-
 static u32 fakeDecStartValue;
 static u64 fakeDecStartTicks;
-static u64 fakeTBStartValue;
-static u64 fakeTBStartTicks;
+
+s64 globalTimer;
+u64 fakeTBStartValue;
+u64 fakeTBStartTicks;
 
 static int ev_lost;
 
diff --git a/Source/Core/Core/CoreTiming.h b/Source/Core/Core/CoreTiming.h
index 15a3a09d89..7eb3784ae1 100644
--- a/Source/Core/Core/CoreTiming.h
+++ b/Source/Core/Core/CoreTiming.h
@@ -25,6 +25,10 @@ class PointerWrap;
 namespace CoreTiming
 {
 
+extern s64 globalTimer;
+extern u64 fakeTBStartValue;
+extern u64 fakeTBStartTicks;
+
 void Init();
 void Shutdown();
 
diff --git a/Source/Core/Core/PowerPC/Jit64/Jit_SystemRegisters.cpp b/Source/Core/Core/PowerPC/Jit64/Jit_SystemRegisters.cpp
index 040a80d465..9c00e70be9 100644
--- a/Source/Core/Core/PowerPC/Jit64/Jit_SystemRegisters.cpp
+++ b/Source/Core/Core/PowerPC/Jit64/Jit_SystemRegisters.cpp
@@ -173,15 +173,25 @@ void Jit64::mfspr(UGeckoInstruction inst)
 		// typical use of this instruction is to call it three times, e.g. mftbu/mftbl/mftbu/cmpw/bne
 		// to deal with possible timer wraparound. This makes the second two (out of three) completely
 		// redundant for the JIT.
-		u32 registersInUse = CallerSavedRegistersInUse();
 		u32 offset = js.downcountAmount / SystemTimers::TIMER_RATIO;
-		ABI_PushRegistersAndAdjustStack(registersInUse, false);
-		ABI_CallFunction((void *)&SystemTimers::GetFakeTimeBase);
-		ABI_PopRegistersAndAdjustStack(registersInUse, false);
+		gpr.FlushLockX(EDX);
+
+		// An inline implementation of CoreTiming::GetFakeTimeBase, since in timer-heavy games the
+		// cost of calling out to C for this is actually significant.
+		MOV(64, R(RAX), M(&CoreTiming::globalTimer));
+		SUB(64, R(RAX), M(&CoreTiming::fakeTBStartTicks));
+		// a / 12 = (a * 0xAAAAAAAAAAAAAAAB) >> 67
+		MOV(64, R(RDX), Imm64(0xAAAAAAAAAAAAAAABULL));
+		MUL(64, R(RDX));
+		MOV(64, R(RAX), M(&CoreTiming::fakeTBStartValue));
+		SHR(64, R(RDX), Imm8(3));
 		// The timer can change within a long block, so add in any difference
 		if (offset > 0)
-			ADD(64, R(RAX), Imm32(offset));
+			LEA(64, RAX, MComplex(RAX, RDX, SCALE_1, offset));
+		else
+			ADD(64, R(RAX), R(RDX));
 		MOV(64, M(&TL), R(RAX));
+
 		// Two calls of TU/TL next to each other are extremely common in typical usage, so merge them
 		// if we can.
 		u32 nextIndex = (js.next_inst.SPRU << 5) | (js.next_inst.SPRL & 0x1F);
@@ -228,6 +238,7 @@ void Jit64::mfspr(UGeckoInstruction inst)
 		break;
 	}
 	gpr.UnlockAll();
+	gpr.UnlockAllX();
 }
 
 void Jit64::mtmsr(UGeckoInstruction inst)