JIT: cleanup instrction merging for idle loop and dcbst.

Use MergeAllowedNextInstructions where appropriate, and read from the
instruction buffer rather than poking directly into memory.

3 files changed, 22 insertions, 13 deletions

diff --git a/Source/Core/Core/PowerPC/Jit64/Jit.h b/Source/Core/Core/PowerPC/Jit64/Jit.h
index a2d6c3026f..2352d90c03 100644
--- a/Source/Core/Core/PowerPC/Jit64/Jit.h
+++ b/Source/Core/Core/PowerPC/Jit64/Jit.h
@@ -241,7 +241,7 @@ public:
 	void negx(UGeckoInstruction inst);
 	void slwx(UGeckoInstruction inst);
 	void srwx(UGeckoInstruction inst);
-	void dcbst(UGeckoInstruction inst);
+	void dcbt(UGeckoInstruction inst);
 	void dcbz(UGeckoInstruction inst);
 
 	void subfic(UGeckoInstruction inst);
diff --git a/Source/Core/Core/PowerPC/Jit64/Jit64_Tables.cpp b/Source/Core/Core/PowerPC/Jit64/Jit64_Tables.cpp
index 636a9b2396..c33f22e8d5 100644
--- a/Source/Core/Core/PowerPC/Jit64/Jit64_Tables.cpp
+++ b/Source/Core/Core/PowerPC/Jit64/Jit64_Tables.cpp
@@ -197,10 +197,10 @@ static GekkoOPTemplate table31[] =
 	{824, &Jit64::srawix},                 //"srawix", OPTYPE_INTEGER, FL_OUT_A | FL_IN_B | FL_IN_S | FL_SET_CA | FL_RC_BIT}},
 	{24,  &Jit64::slwx},                   //"slwx",   OPTYPE_INTEGER, FL_OUT_A | FL_IN_B | FL_IN_S | FL_RC_BIT}},
 
-	{54,   &Jit64::dcbst},                 //"dcbst",  OPTYPE_DCACHE, 0, 4}},
+	{54,   &Jit64::FallBackToInterpreter}, //"dcbst",  OPTYPE_DCACHE, 0, 4}},
 	{86,   &Jit64::FallBackToInterpreter}, //"dcbf",   OPTYPE_DCACHE, 0, 4}},
-	{246,  &Jit64::DoNothing},             //"dcbtst", OPTYPE_DCACHE, 0, 1}},
-	{278,  &Jit64::DoNothing},             //"dcbt",   OPTYPE_DCACHE, 0, 1}},
+	{246,  &Jit64::dcbt },                 //"dcbtst", OPTYPE_DCACHE, 0, 1}},
+	{278,  &Jit64::dcbt },                 //"dcbt",   OPTYPE_DCACHE, 0, 1}},
 	{470,  &Jit64::FallBackToInterpreter}, //"dcbi",   OPTYPE_DCACHE, 0, 4}},
 	{758,  &Jit64::DoNothing},             //"dcba",   OPTYPE_DCACHE, 0, 4}},
 	{1014, &Jit64::dcbz},                  //"dcbz",   OPTYPE_DCACHE, 0, 4}},
diff --git a/Source/Core/Core/PowerPC/Jit64/Jit_LoadStore.cpp b/Source/Core/Core/PowerPC/Jit64/Jit_LoadStore.cpp
index 0c0a23acfc..29e83e034d 100644
--- a/Source/Core/Core/PowerPC/Jit64/Jit_LoadStore.cpp
+++ b/Source/Core/Core/PowerPC/Jit64/Jit_LoadStore.cpp
@@ -111,10 +111,11 @@ void Jit64::lXXx(UGeckoInstruction inst)
 	if (SConfig::GetInstance().m_LocalCoreStartupParameter.bSkipIdle &&
 	    PowerPC::GetState() != PowerPC::CPU_STEPPING &&
 	    inst.OPCD == 32 &&
+	    MergeAllowedNextInstructions(2) &&
 	    (inst.hex & 0xFFFF0000) == 0x800D0000 &&
-	    (Memory::ReadUnchecked_U32(js.compilerPC + 4) == 0x28000000 ||
-	    (SConfig::GetInstance().m_LocalCoreStartupParameter.bWii && Memory::ReadUnchecked_U32(js.compilerPC + 4) == 0x2C000000)) &&
-	    Memory::ReadUnchecked_U32(js.compilerPC + 8) == 0x4182fff8)
+	    (js.op[1].inst.hex == 0x28000000 ||
+	    (SConfig::GetInstance().m_LocalCoreStartupParameter.bWii && js.op[1].inst.hex == 0x2C000000)) &&
+	    js.op[2].inst.hex == 0x4182fff8)
 	{
 		// TODO(LinesPrower):
 		// - Rewrite this!
@@ -284,16 +285,24 @@ void Jit64::lXXx(UGeckoInstruction inst)
 	gpr.UnlockAllX();
 }
 
-void Jit64::dcbst(UGeckoInstruction inst)
+void Jit64::dcbt(UGeckoInstruction inst)
 {
 	INSTRUCTION_START
 	JITDISABLE(bJITLoadStoreOff);
 
-	// If the dcbst instruction is preceded by dcbt, it is flushing a prefetched
-	// memory location.  Do not invalidate the JIT cache in this case as the memory
-	// will be the same.
-	// dcbt = 0x7c00022c
-	FALLBACK_IF((Memory::ReadUnchecked_U32(js.compilerPC - 4) & 0x7c00022c) != 0x7c00022c);
+	// Prefetch. Since we don't emulate the data cache, we don't need to do anything.
+
+	// If a dcbst follows a dcbt, it probably isn't a case of dynamic code
+	// modification, so don't bother invalidating the jit block cache.
+	// This is important because invalidating the block cache when we don't
+	// need to is terrible for performance.
+	// (Invalidating the jit block cache on dcbst is a heuristic.)
+	if (MergeAllowedNextInstructions(1) &&
+	    js.op[1].inst.OPCD == 31 && js.op[1].inst.SUBOP10 == 54 &&
+	    js.op[1].inst.RA == inst.RA && js.op[1].inst.RB == inst.RB)
+	{
+		js.skipInstructions = 1;
+	}
 }
 
 // Zero cache line.