mirror of https://github.com/PCSX2/pcsx2.git
*experimental* Added forced block splitting at 4k code boundaries in PS2 memory. This fixes issues with frequent recompilation of certain uncounted manual blocks, and also allows for us to greatly simplify the memory protection and recompiler block management code, but *may* have performance penalties under specific conditions. So I've added a verbose log message to Devel builds so that we can look for any problem games that might invoke more than their fair share of tiny block splits.
git-svn-id: http://pcsx2.googlecode.com/svn/trunk@1093 96395faa-99c1-11dd-bbfe-3dabce05a288
This commit is contained in:
Jake.Stine
parent
3af9acc01d
commit
d14259b8b4
|
|
@ -1090,10 +1090,8 @@ void recompileNextInstruction(int delayslot)
|
|||
s_pCode = (int *)PSM( pc );
|
||||
assert(s_pCode);
|
||||
|
||||
// why?
|
||||
#ifdef _DEBUG
|
||||
MOV32ItoR(EAX, pc);
|
||||
#endif
|
||||
if( IsDebugBuild )
|
||||
MOV32ItoR(EAX, pc); // acts as a tag for delimiting recompiled instructions when viewing x86 disasm.
|
||||
|
||||
cpuRegs.code = *(int *)s_pCode;
|
||||
pc += 4;
|
||||
|
|
@ -1142,63 +1140,33 @@ void recompileNextInstruction(int delayslot)
|
|||
|
||||
const OPCODE& opcode = GetCurrentInstruction();
|
||||
|
||||
// peephole optimizations
|
||||
#ifdef PCSX2_VM_COISSUE
|
||||
if( g_pCurInstInfo->info & EEINSTINFO_COREC ) {
|
||||
//assert( !(g_pCurInstInfo->info & EEINSTINFO_NOREC) );
|
||||
|
||||
if( g_pCurInstInfo->numpeeps > 1 ) {
|
||||
switch(_Opcode_) {
|
||||
case 30: recLQ_coX(g_pCurInstInfo->numpeeps); break;
|
||||
case 31: recSQ_coX(g_pCurInstInfo->numpeeps); break;
|
||||
case 49: recLWC1_coX(g_pCurInstInfo->numpeeps); break;
|
||||
case 57: recSWC1_coX(g_pCurInstInfo->numpeeps); break;
|
||||
case 55: recLD_coX(g_pCurInstInfo->numpeeps); break;
|
||||
case 63: recSD_coX(g_pCurInstInfo->numpeeps, 1); break; //not sure if should be set to 1 or 0; looks like "1" handles alignment, so i'm going with that for now
|
||||
// if this instruction is a jump or a branch, exit right away
|
||||
if( delayslot ) {
|
||||
switch(_Opcode_) {
|
||||
case 1:
|
||||
switch(_Rt_) {
|
||||
case 0: case 1: case 2: case 3: case 0x10: case 0x11: case 0x12: case 0x13:
|
||||
Console::Notice("branch %x in delay slot!", params cpuRegs.code);
|
||||
_clearNeededX86regs();
|
||||
_clearNeededMMXregs();
|
||||
_clearNeededXMMregs();
|
||||
return;
|
||||
}
|
||||
break;
|
||||
|
||||
jNO_DEFAULT
|
||||
}
|
||||
pc += g_pCurInstInfo->numpeeps*4;
|
||||
s_nBlockCycles += (g_pCurInstInfo->numpeeps+1) * opcode.cycles;
|
||||
g_pCurInstInfo += g_pCurInstInfo->numpeeps;
|
||||
}
|
||||
else {
|
||||
recBSC_co[_Opcode_]();
|
||||
pc += 4;
|
||||
g_pCurInstInfo++;
|
||||
s_nBlockCycles += opcode.cycles*2;
|
||||
case 2: case 3: case 4: case 5: case 6: case 7: case 0x14: case 0x15: case 0x16: case 0x17:
|
||||
Console::Notice("branch %x in delay slot!", params cpuRegs.code);
|
||||
_clearNeededX86regs();
|
||||
_clearNeededMMXregs();
|
||||
_clearNeededXMMregs();
|
||||
return;
|
||||
}
|
||||
}
|
||||
else
|
||||
#endif
|
||||
{
|
||||
//assert( !(g_pCurInstInfo->info & EEINSTINFO_NOREC) );
|
||||
|
||||
// if this instruction is a jump or a branch, exit right away
|
||||
if( delayslot ) {
|
||||
switch(_Opcode_) {
|
||||
case 1:
|
||||
switch(_Rt_) {
|
||||
case 0: case 1: case 2: case 3: case 0x10: case 0x11: case 0x12: case 0x13:
|
||||
Console::Notice("branch %x in delay slot!", params cpuRegs.code);
|
||||
_clearNeededX86regs();
|
||||
_clearNeededMMXregs();
|
||||
_clearNeededXMMregs();
|
||||
return;
|
||||
}
|
||||
break;
|
||||
|
||||
case 2: case 3: case 4: case 5: case 6: case 7: case 0x14: case 0x15: case 0x16: case 0x17:
|
||||
Console::Notice("branch %x in delay slot!", params cpuRegs.code);
|
||||
_clearNeededX86regs();
|
||||
_clearNeededMMXregs();
|
||||
_clearNeededXMMregs();
|
||||
return;
|
||||
}
|
||||
}
|
||||
//If thh COP0 DIE bit is disabled, double the cycles. Happens rarely.
|
||||
s_nBlockCycles += opcode.cycles * (2 - ((cpuRegs.CP0.n.Config >> 18) & 0x1));
|
||||
opcode.recompile();
|
||||
}
|
||||
//If thh COP0 DIE bit is disabled, double the cycles. Happens rarely.
|
||||
s_nBlockCycles += opcode.cycles * (2 - ((cpuRegs.CP0.n.Config >> 18) & 0x1));
|
||||
opcode.recompile();
|
||||
|
||||
if( !delayslot ) {
|
||||
if( s_bFlushReg ) {
|
||||
|
|
@ -1358,25 +1326,41 @@ void recRecompile( const u32 startpc )
|
|||
|
||||
while(1) {
|
||||
BASEBLOCK* pblock = PC_GETBLOCK(i);
|
||||
if (i != startpc && pblock->GetFnptr() != (uptr)JITCompile && pblock->GetFnptr() != (uptr)JITCompileInBlock) {
|
||||
// branch = 3
|
||||
willbranch3 = 1;
|
||||
s_nEndBlock = i;
|
||||
break;
|
||||
|
||||
if(i != startpc) // Block size truncation checks.
|
||||
{
|
||||
if( (i & 0xffc) == 0x0 ) // breaks blocks at 4k page boundaries
|
||||
{
|
||||
willbranch3 = 1;
|
||||
s_nEndBlock = i;
|
||||
|
||||
// Log the pagesplits verbosely for now, until we see if any games are affected
|
||||
// adversely by excessive splits.
|
||||
DevCon::Notice( "Pagesplit @ %08X : size=%d insts", params startpc, (i-startpc) / 4 );
|
||||
|
||||
break;
|
||||
}
|
||||
|
||||
if (pblock->GetFnptr() != (uptr)JITCompile && pblock->GetFnptr() != (uptr)JITCompileInBlock)
|
||||
{
|
||||
willbranch3 = 1;
|
||||
s_nEndBlock = i;
|
||||
break;
|
||||
}
|
||||
}
|
||||
|
||||
//HUH ? PSM ? whut ? THIS IS VIRTUAL ACCESS GOD DAMMIT
|
||||
cpuRegs.code = *(int *)PSM(i);
|
||||
|
||||
switch(cpuRegs.code >> 26) {
|
||||
case 0: // special
|
||||
|
||||
if( _Funct_ == 8 || _Funct_ == 9 ) { // JR, JALR
|
||||
s_nEndBlock = i + 8;
|
||||
s_nHasDelay = 1;
|
||||
goto StartRecomp;
|
||||
}
|
||||
|
||||
break;
|
||||
|
||||
case 1: // regimm
|
||||
|
||||
if( _Rt_ < 4 || (_Rt_ >= 16 && _Rt_ < 20) ) {
|
||||
|
|
@ -1390,7 +1374,6 @@ void recRecompile( const u32 startpc )
|
|||
|
||||
goto StartRecomp;
|
||||
}
|
||||
|
||||
break;
|
||||
|
||||
case 2: // J
|
||||
|
|
@ -1525,8 +1508,11 @@ StartRecomp:
|
|||
iDumpBlock(startpc, recPtr);
|
||||
#endif
|
||||
|
||||
u32 sz = (s_nEndBlock-startpc)>>2;
|
||||
// fixme! The following manual/protected block code can be greatly simplified now.
|
||||
// It originally had to account for cross-page blocks, but we have since guaranteed
|
||||
// that no block will cross a page boundary.
|
||||
|
||||
u32 sz = (s_nEndBlock-startpc) >> 2;
|
||||
u32 inpage_ptr = HWADDR(startpc);
|
||||
u32 inpage_sz = sz*4;
|
||||
|
||||
|
|
@ -1559,15 +1545,34 @@ StartRecomp:
|
|||
lpc += 4;
|
||||
}
|
||||
|
||||
if (startpc != 0x81fc0 && manual_counter[inpage_ptr >> 12] <= 4) {
|
||||
xADD(ptr16[&manual_page[inpage_ptr >> 12]], 1);
|
||||
xJC( dyna_page_reset );
|
||||
// Tweakpoint! 3 is a 'magic' number representing the number of times a counted block
|
||||
// is re-protected before the recompiler gives up and sets it up as an uncounted (permanent)
|
||||
// manual block. 4 definitely seemed too high, but 2 might be better? Side effects of a
|
||||
// lower threshold: over extended gameplay with several map changes, a game's overall
|
||||
// performance could degrade.
|
||||
|
||||
// KH2 manual_counter failure -- during the intro vid a block which overlaps
|
||||
// two pages causes constant recompilation. The block is very large and has two entry
|
||||
// points, one around 0x01f1de1c and one around 0x01f1dfac (varies on locale). The
|
||||
// former is the one that's not being cleared correctly [or at least behaves as though
|
||||
// it's not being cleared correctly].
|
||||
// (ideally, perhaps, manual_counter should be reset to 0 every few minutes?)
|
||||
|
||||
if (startpc != 0x81fc0 && manual_counter[inpage_ptr >> 12] <= 3) {
|
||||
|
||||
// Counted blocks add a weighted (by block size) value into manual_page each time they're
|
||||
// run. If the block gets run a lot, it resets and re-protects itself in the hope
|
||||
// that whatever forced it to be manually-checked before was a 1-time deal.
|
||||
|
||||
// Counted blocks have a secondary threshold check in manual_counter, which forces a block
|
||||
// to 'uncounted' mode if it's recompiled several time. This protects against excessive
|
||||
// recompilation of blocks that reside on the same codepage as data.
|
||||
|
||||
// fixme? Currently this algo is kinda dumb and results in the forced recompilation of a
|
||||
// lot of blocks before it decides to mark a 'busy' page as uncounted. There might be
|
||||
// be a more clever approach that could streamline this process, by doing a first-pass
|
||||
// test using the vtlb memory protection (without recompilation!) to reprotect a counted
|
||||
// block. But unless a new also is relatively simple in implementation, it's probably
|
||||
// not worth the effort (tests show that we have lots of recompiler memory to spare, and
|
||||
// that the current amount of recompilation is fairly cheap).
|
||||
|
||||
xADD(ptr16[&manual_page[inpage_ptr >> 12]], sz);
|
||||
xJC( dyna_page_reset );
|
||||
|
||||
// note: clearcnt is measured per-page, not per-block!
|
||||
DbgCon::WriteLn( "Manual block @ %08X : size=%3d page/offs=%05X/%03X inpgsz=%d clearcnt=%d",
|
||||
|
|
@ -1575,7 +1580,7 @@ StartRecomp:
|
|||
}
|
||||
else
|
||||
{
|
||||
DbgCon::WriteLn( "Uncounted Manual block @ %08X : size=%3d page/offs=%05X/%03X inpgsz=%d",
|
||||
DbgCon::Notice( "Uncounted Manual block @ %08X : size=%3d page/offs=%05X/%03X inpgsz=%d",
|
||||
params startpc, sz, inpage_ptr>>12, inpage_offs, pgsz, inpage_sz );
|
||||
}
|
||||
|
||||
|
|
@ -1584,11 +1589,11 @@ StartRecomp:
|
|||
inpage_ptr += pgsz;
|
||||
inpage_sz -= pgsz;
|
||||
}
|
||||
|
||||
// finally recompile //
|
||||
|
||||
// Finally: Generate x86 recompiled code!
|
||||
g_pCurInstInfo = s_pInstCache;
|
||||
while (!branch && pc < s_nEndBlock) {
|
||||
recompileNextInstruction(0);
|
||||
recompileNextInstruction(0); // For the love of recursion, batman!
|
||||
}
|
||||
|
||||
#ifdef _DEBUG
|
||||
|
|
@ -1653,8 +1658,20 @@ StartRecomp:
|
|||
ADD32ItoM((int)&cpuRegs.cycle, eeScaleBlockCycles() );
|
||||
|
||||
if( willbranch3 || !branch) {
|
||||
|
||||
iFlushCall(FLUSH_EVERYTHING);
|
||||
iBranchTest(pc);
|
||||
|
||||
// Split Block concatenation mode.
|
||||
// This code is run when blocks are split either to keep block sizes manageable
|
||||
// or because we're crossing a 4k page protection boundary in ps2 mem. The latter
|
||||
// case can result in very short blocks which should not issue branch tests for
|
||||
// performance reasons.
|
||||
|
||||
int numinsts = (pc - startpc) / 4;
|
||||
if( numinsts > 12 )
|
||||
iBranchTest(pc);
|
||||
else
|
||||
iBranch(pc,0); // unconditional static link
|
||||
}
|
||||
}
|
||||
|
||||
|
|
|
|||
Loading…
Reference in New Issue