Fixed bugs in BTS instruction emitters, and working on fixing bugs in BTS memory protection.

git-svn-id: http://pcsx2.googlecode.com/svn/trunk@1089 96395faa-99c1-11dd-bbfe-3dabce05a288

pcsx2/x86/ix86-32/iR5900-32.cpp

@@ -751,6 +751,13 @@ void recClear(u32 addr, u32 size)
 		return;
 	addr = HWADDR(addr);
 
+	addr -= addr % 16;		// round down.
+	size += 3;				// round up!
+	size -= size % 4;
+
+	for (u32 a = addr / 16; a < addr / 16 + size / 4; a++)
+		vtlb_private::vtlbdata.alloc_bits[a / 8] &= ~(1 << (a & 7));
+
 	int blockidx = recBlocks.LastIndex(addr + size * 4 - 4);
 
 	if (blockidx == -1)
@@ -1544,66 +1551,32 @@ StartRecomp:
 			}
 			else
 			{
-				// import the vtlbdata (alloc_bits and alloc_base and stuff):
 				using namespace vtlb_private;
-				
+
 				MOV32ItoR(ECX, inpage_ptr);
-				MOV32ItoR(EDX, pgsz);
+				MOV32ItoR(EDX, pgsz / 4);
-				
-				u32 mask=0;
-				u32 writen=0;
-				u32 writen_start=0;
 
-				u32 lpc=inpage_ptr;
+				u32 index = (psM - vtlbdata.alloc_base + inpage_ptr) / 16 / 32;		// 16 bytes per bit, 32 bits per dword.
-				u32 stg=pgsz;
+				u32 mask = 0;
 
-				while(stg>0)
+				u32 start = inpage_ptr & ~15;
-				{
+				u32 end = inpage_ptr + pgsz;
-					u32 bit = (lpc>>4) & 7;
+				for (u32 pos = start, bit = (start / 16) % 32; pos < end; pos += 16, bit++) {
-					if (mask==0)
+					if( bit == 32 )
 					{
-						//writen=bit;
+						xTEST(ptr32[&vtlbdata.alloc_bits[index]], mask);
-						writen_start=(((u8*)PSM(lpc)-vtlbdata.alloc_base)>>4)/8;
+						xJNZ(dyna_block_discard);
+						bit = 0;
+						mask = 0;
+						index++;
 					}
 					mask |= 1 << bit;
-
-					if (bit==31)
-					{
-						vtlbdata.alloc_bits[writen_start]&=~mask;
-						xTEST( ptr32[&vtlbdata.alloc_bits[writen_start]], mask );	// auto-optimizes to imm8 when applicable.
-						xJNZ( dyna_block_discard );
-						//SysPrintf("%08X %d %d\n",mask,pgsz,pgsz>>4);
-						mask = 0;
-					}
-
-					//writen++;
-
-					if (stg<=16)
-					{
-						lpc += stg;
-						stg  = 0;
-					}
-					else
-					{
-						lpc += 16;
-						stg -= 16;
-					}
 				}
+				xTEST(ptr32[&vtlbdata.alloc_bits[index]], mask);
+				xJNZ(dyna_block_discard);
 
-				if (mask)
+				if (manual_counter[inpage_ptr >> 12] <= 4)
 				{
-					vtlbdata.alloc_bits[writen_start] &= ~mask;
-					xTEST( ptr32[&vtlbdata.alloc_bits[writen_start]], mask );	// auto-optimizes to imm8 when applicable.
-					xJNZ( dyna_block_discard );
-					//SysPrintf("%08X %d %d\n",mask,pgsz,pgsz>>4);
-					mask = 0;
-				}
-
-				if( startpc != 0x81fc0 && manual_counter[inpage_ptr >> 12] <= 4 )
-				{
-					// Commented out until we replace it with a smarter algo that only
-					// recompiles blocks a limited number of times.
-
 					xADD(ptr16[&manual_page[inpage_ptr >> 12]], 1);
 					xJC( dyna_page_reset );
 				}

pcsx2/x86/ix86-32/recVTLB.cpp

@@ -430,7 +430,7 @@ static void _vtlb_DynGen_DirectWrite( u32 bits )
 	u8* bits_base = vtlbdata.alloc_bits;
 	bits_base -= (alloc_base>>4)/8;		//in bytes
 
-	xBTS( ecx, bits_base );
+	xBTS( ptr32[bits_base], ecx );
 }
 
 // ------------------------------------------------------------------------

pcsx2/x86/ix86/implement/test.h

@@ -96,15 +96,23 @@ public:
 // Bit Test Instructions - Valid on 16/32 bit instructions only.
 //
 template< G8Type InstType >
-class xImpl_Group8 : public xImpl_BitScan<0xa3 | (InstType << 2)>
+class xImpl_Group8
 {
+	static const uint RegFormOp = 0xa3 | (InstType << 3);
 public:
-	using xImpl_BitScan<0xa3 | (InstType << 2)>::operator();
+	__forceinline void operator()( const xRegister32& bitbase, const xRegister32& bitoffset ) const	{ xOpWrite0F( RegFormOp, bitbase, bitoffset ); }
+	__forceinline void operator()( const xRegister16& bitbase, const xRegister16& bitoffset ) const	{ xOpWrite0F( 0x66, RegFormOp, bitbase, bitoffset ); }
+	__forceinline void operator()( const ModSibBase& bitbase, const xRegister32& bitoffset ) const	{ xOpWrite0F( RegFormOp, bitoffset, bitbase ); }
+	__forceinline void operator()( const ModSibBase& bitbase, const xRegister16& bitoffset ) const	{ xOpWrite0F( 0x66, RegFormOp, bitoffset, bitbase ); }
+	__forceinline void operator()( const void* bitbase, const xRegister32& bitoffset ) const		{ xOpWrite0F( 0xab, bitoffset, bitbase ); }
+	__forceinline void operator()( const void* bitbase, const xRegister16& bitoffset ) const		{ xOpWrite0F( 0x66, RegFormOp, bitoffset, bitbase ); }
 
-	__forceinline void operator()( const ModSibStrict<u32>& bitbase, u8 bitoffset ) const	{ xOpWrite0F( 0xba, InstType, bitbase, bitoffset ); }
+	__forceinline void operator()( const ModSibStrict<u32>& bitbase, u8 bitoffset ) const			{ xOpWrite0F( 0xba, InstType, bitbase, bitoffset ); }
-	__forceinline void operator()( const ModSibStrict<u16>& bitbase, u8 bitoffset ) const	{ xOpWrite0F( 0x66, 0xba, InstType, bitbase, bitoffset ); }
+	__forceinline void operator()( const ModSibStrict<u16>& bitbase, u8 bitoffset ) const			{ xOpWrite0F( 0x66, 0xba, InstType, bitbase, bitoffset ); }
-	void operator()( const xRegister<u32>& bitbase, u8 bitoffset ) const					{ xOpWrite0F( 0xba, InstType, bitbase, bitoffset ); }
+	__forceinline void operator()( const u32* bitbase, u8 bitoffset ) const							{ xOpWrite0F( 0xba, InstType, bitbase, bitoffset ); }
-	void operator()( const xRegister<u16>& bitbase, u8 bitoffset ) const					{ xOpWrite0F( 0x66, 0xba, InstType, bitbase, bitoffset ); }
+	__forceinline void operator()( const u16* bitbase, u8 bitoffset ) const							{ xOpWrite0F( 0x66, 0xba, InstType, bitbase, bitoffset ); }
+	__forceinline void operator()( const xRegister<u32>& bitbase, u8 bitoffset ) const				{ xOpWrite0F( 0xba, InstType, bitbase, bitoffset ); }
+	__forceinline void operator()( const xRegister<u16>& bitbase, u8 bitoffset ) const				{ xOpWrite0F( 0x66, 0xba, InstType, bitbase, bitoffset ); }
 
 	xImpl_Group8() {}
 };

pcsx2/x86/ix86/ix86_tools.cpp

@@ -34,12 +34,12 @@ namespace x86Emitter
 {
 	void xStoreReg( const xRegisterSSE& src )
 	{
-		xMOVDQA( &g_globalXMMData[src.Id], src );
+		xMOVDQA( &g_globalXMMData[src.Id*2], src );
 	}
 
 	void xRestoreReg( const xRegisterSSE& dest )
 	{
-		xMOVDQA( dest, &g_globalXMMData[dest.Id] );
+		xMOVDQA( dest, &g_globalXMMData[dest.Id*2] );
 	}
 }