izzy/dolphin
0
0
Fork 0
mirror of https://github.com/izzy2lost/dolphin.git synced 2026-06-19 01:16:48 -07:00
Code Issues Packages Projects Releases Wiki Activity

Merge pull request #323 from Sonicadvance1/newer-ppcanalyst

Browse Source 
[RFC] New PPCAnalyst class.
This commit is contained in:
Ryan Houdek
2014-04-30 18:59:24 -05:00
parent 1b581b7c27 8e1dfef14c
commit c4221e8f68
13 changed files with 522 additions and 465 deletions
Download Patch File Download Diff File Expand all files Collapse all files
Source/Core/Core/PowerPC/Jit64/Jit.cpp
+15 -25
View File
@@ -181,6 +181,11 @@ void Jit64::Init()
blocks.Init();
asm_routines.Init();
code_block.m_stats = &js.st;
code_block.m_gpa = &js.gpa;
code_block.m_fpa = &js.fpa;
analyzer.SetOption(PPCAnalyst::PPCAnalyzer::OPTION_CONDITIONAL_CONTINUE);
}
void Jit64::ClearCache()
@@ -404,9 +409,6 @@ const u8* Jit64::DoJit(u32 em_address, PPCAnalyst::CodeBuffer *code_buf, JitBloc
// Memory exception on instruction fetch
bool memory_exception = false;
// A broken block is a block that does not end in a branch
bool broken_block = false;
if (Core::g_CoreStartupParameter.bEnableDebugging)
{
// Comment out the following to disable breakpoints (speed-up)
@@ -433,7 +435,6 @@ const u8* Jit64::DoJit(u32 em_address, PPCAnalyst::CodeBuffer *code_buf, JitBloc
}
}
int size = 0;
js.firstFPInstructionFound = false;
js.isLastInstruction = false;
js.blockStart = em_address;
@@ -444,17 +445,12 @@ const u8* Jit64::DoJit(u32 em_address, PPCAnalyst::CodeBuffer *code_buf, JitBloc
jit->js.numLoadStoreInst = 0;
jit->js.numFloatingPointInst = 0;
u32 nextPC = em_address;
// Analyze the block, collect all instructions it is made of (including inlining,
// if that is enabled), reorder instructions for optimal performance, and join joinable instructions.
u32 nextPC = em_address;
u32 merged_addresses[32];
const int capacity_of_merged_addresses = sizeof(merged_addresses) / sizeof(merged_addresses[0]);
int size_of_merged_addresses = 0;
if (!memory_exception)
{
// If there is a memory exception inside a block (broken_block==true), compile up to that instruction.
nextPC = PPCAnalyst::Flatten(em_address, &size, &js.st, &js.gpa, &js.fpa, broken_block, code_buf, blockSize, merged_addresses, capacity_of_merged_addresses, size_of_merged_addresses);
}
nextPC = analyzer.Analyze(em_address, &code_block, code_buf, blockSize);
PPCAnalyst::CodeOp *ops = code_buf->codebuffer;
@@ -499,19 +495,13 @@ const u8* Jit64::DoJit(u32 em_address, PPCAnalyst::CodeBuffer *code_buf, JitBloc
js.downcountAmount = 0;
if (!Core::g_CoreStartupParameter.bEnableDebugging)
{
for (int i = 0; i < size_of_merged_addresses; ++i)
{
const u32 address = merged_addresses[i];
js.downcountAmount += PatchEngine::GetSpeedhackCycles(address);
}
}
js.downcountAmount += PatchEngine::GetSpeedhackCycles(code_block.m_address);
js.skipnext = false;
js.blockSize = size;
js.blockSize = code_block.m_num_instructions;
js.compilerPC = nextPC;
// Translate instructions
for (int i = 0; i < (int)size; i++)
for (u32 i = 0; i < code_block.m_num_instructions; i++)
{
js.compilerPC = ops[i].address;
js.op = &ops[i];
@@ -519,7 +509,7 @@ const u8* Jit64::DoJit(u32 em_address, PPCAnalyst::CodeBuffer *code_buf, JitBloc
const GekkoOPInfo *opinfo = ops[i].opinfo;
js.downcountAmount += opinfo->numCycles;
if (i == (int)size - 1)
if (i == (code_block.m_num_instructions - 1))
{
// WARNING - cmp->branch merging will screw this up.
js.isLastInstruction = true;
@@ -706,7 +696,7 @@ const u8* Jit64::DoJit(u32 em_address, PPCAnalyst::CodeBuffer *code_buf, JitBloc
WriteExceptionExit();
}
if (broken_block)
if (code_block.m_broken)
{
gpr.Flush(FLUSH_ALL);
fpr.Flush(FLUSH_ALL);
@@ -715,10 +705,10 @@ const u8* Jit64::DoJit(u32 em_address, PPCAnalyst::CodeBuffer *code_buf, JitBloc
b->flags = js.block_flags;
b->codeSize = (u32)(GetCodePtr() - normalEntry);
b->originalSize = size;
b->originalSize = code_block.m_num_instructions;
#ifdef JIT_LOG_X86
LogGeneratedX86(size, code_buf, normalEntry, b);
LogGeneratedX86(code_block.m_num_instructions, code_buf, normalEntry, b);
#endif
return normalEntry;
Source/Core/Core/PowerPC/Jit64/Jit_Branch.cpp
+9 -11
View File
@@ -103,7 +103,6 @@ void Jit64::bcx(UGeckoInstruction inst)
JITDISABLE(bJITBranchOff)
// USES_CR
_assert_msg_(DYNA_REC, js.isLastInstruction, "bcx not last instruction of block");
gpr.Flush(FLUSH_ALL);
fpr.Flush(FLUSH_ALL);
@@ -142,7 +141,9 @@ void Jit64::bcx(UGeckoInstruction inst)
SetJumpTarget( pConditionDontBranch );
if ((inst.BO & BO_DONT_DECREMENT_FLAG) == 0)
SetJumpTarget( pCTRDontBranch );
WriteExit(js.compilerPC + 4);
if (!analyzer.HasOption(PPCAnalyst::PPCAnalyzer::OPTION_CONDITIONAL_CONTINUE))
WriteExit(js.compilerPC + 4);
}
void Jit64::bcctrx(UGeckoInstruction inst)
@@ -190,7 +191,9 @@ void Jit64::bcctrx(UGeckoInstruction inst)
WriteExitDestInEAX();
// Would really like to continue the block here, but it ends. TODO.
SetJumpTarget(b);
WriteExit(js.compilerPC + 4);
if (!analyzer.HasOption(PPCAnalyst::PPCAnalyzer::OPTION_CONDITIONAL_CONTINUE))
WriteExit(js.compilerPC + 4);
}
}
@@ -199,13 +202,6 @@ void Jit64::bclrx(UGeckoInstruction inst)
INSTRUCTION_START
JITDISABLE(bJITBranchOff)
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);
@@ -245,5 +241,7 @@ void Jit64::bclrx(UGeckoInstruction inst)
SetJumpTarget( pConditionDontBranch );
if ((inst.BO & BO_DONT_DECREMENT_FLAG) == 0)
SetJumpTarget( pCTRDontBranch );
WriteExit(js.compilerPC + 4);
if (!analyzer.HasOption(PPCAnalyst::PPCAnalyzer::OPTION_CONDITIONAL_CONTINUE))
WriteExit(js.compilerPC + 4);
}
Source/Core/Core/PowerPC/Jit64/Jit_Integer.cpp
+13 -7
View File
@@ -425,10 +425,12 @@ void Jit64::cmpXX(UGeckoInstruction inst)
}
else
{
WriteExit(js.next_compilerPC + 4);
if (!analyzer.HasOption(PPCAnalyst::PPCAnalyzer::OPTION_CONDITIONAL_CONTINUE))
{
js.skipnext = true;
WriteExit(js.next_compilerPC + 4);
}
}
js.cancel = true;
}
}
else
@@ -535,9 +537,11 @@ void Jit64::cmpXX(UGeckoInstruction inst)
if (!!(4 & test_bit) == condition) SetJumpTarget(continue2);
if (!!(2 & test_bit) == condition) SetJumpTarget(continue1);
WriteExit(js.next_compilerPC + 4);
js.cancel = true;
if (!analyzer.HasOption(PPCAnalyst::PPCAnalyzer::OPTION_CONDITIONAL_CONTINUE))
{
js.skipnext = true;
WriteExit(js.next_compilerPC + 4);
}
}
}
@@ -2229,5 +2233,7 @@ void Jit64::twx(UGeckoInstruction inst)
SetJumpTarget(exit3);
SetJumpTarget(exit4);
SetJumpTarget(exit5);
WriteExit(js.compilerPC + 4);
if (!analyzer.HasOption(PPCAnalyst::PPCAnalyzer::OPTION_CONDITIONAL_CONTINUE))
WriteExit(js.compilerPC + 4);
}
Source/Core/Core/PowerPC/Jit64IL/JitIL.cpp
+14 -27
View File
@@ -268,6 +268,10 @@ void JitIL::Init()
blocks.Init();
asm_routines.Init();
code_block.m_stats = &js.st;
code_block.m_gpa = &js.gpa;
code_block.m_fpa = &js.fpa;
if (SConfig::GetInstance().m_LocalCoreStartupParameter.bJITILTimeProfiling) {
JitILProfiler::Init();
}
@@ -500,9 +504,6 @@ const u8* JitIL::DoJit(u32 em_address, PPCAnalyst::CodeBuffer *code_buf, JitBloc
// Memory exception on instruction fetch
bool memory_exception = false;
// A broken block is a block that does not end in a branch
bool broken_block = false;
if (Core::g_CoreStartupParameter.bEnableDebugging)
{
// Comment out the following to disable breakpoints (speed-up)
@@ -529,7 +530,6 @@ const u8* JitIL::DoJit(u32 em_address, PPCAnalyst::CodeBuffer *code_buf, JitBloc
}
}
int size = 0;
js.isLastInstruction = false;
js.blockStart = em_address;
js.fifoBytesThisBlock = 0;
@@ -538,19 +538,12 @@ const u8* JitIL::DoJit(u32 em_address, PPCAnalyst::CodeBuffer *code_buf, JitBloc
jit->js.numLoadStoreInst = 0;
jit->js.numFloatingPointInst = 0;
u32 nextPC = em_address;
// Analyze the block, collect all instructions it is made of (including inlining,
// if that is enabled), reorder instructions for optimal performance, and join joinable instructions.
u32 exitAddress = em_address;
u32 merged_addresses[32];
const int capacity_of_merged_addresses = sizeof(merged_addresses) / sizeof(merged_addresses[0]);
int size_of_merged_addresses = 0;
if (!memory_exception)
{
// If there is a memory exception inside a block (broken_block==true), compile up to that instruction.
// TODO
exitAddress = PPCAnalyst::Flatten(em_address, &size, &js.st, &js.gpa, &js.fpa, broken_block, code_buf, blockSize, merged_addresses, capacity_of_merged_addresses, size_of_merged_addresses);
}
nextPC = analyzer.Analyze(em_address, &code_block, code_buf, blockSize);
PPCAnalyst::CodeOp *ops = code_buf->codebuffer;
const u8 *start = AlignCode4(); // TODO: Test if this or AlignCode16 make a difference from GetCodePtr
@@ -586,7 +579,7 @@ const u8* JitIL::DoJit(u32 em_address, PPCAnalyst::CodeBuffer *code_buf, JitBloc
SConfig::GetInstance().m_LocalCoreStartupParameter.bJITILOutputIR)
{
// For profiling and IR Writer
for (int i = 0; i < (int)size; i++)
for (u32 i = 0; i < code_block.m_num_instructions; i++)
{
const u64 inst = ops[i].inst.hex;
// Ported from boost::hash
@@ -606,16 +599,10 @@ const u8* JitIL::DoJit(u32 em_address, PPCAnalyst::CodeBuffer *code_buf, JitBloc
js.downcountAmount = 0;
if (!Core::g_CoreStartupParameter.bEnableDebugging)
{
for (int i = 0; i < size_of_merged_addresses; ++i)
{
const u32 address = merged_addresses[i];
js.downcountAmount += PatchEngine::GetSpeedhackCycles(address);
}
}
js.downcountAmount += PatchEngine::GetSpeedhackCycles(code_block.m_address);
// Translate instructions
for (int i = 0; i < (int)size; i++)
for (u32 i = 0; i < code_block.m_num_instructions; i++)
{
js.compilerPC = ops[i].address;
js.op = &ops[i];
@@ -623,7 +610,7 @@ const u8* JitIL::DoJit(u32 em_address, PPCAnalyst::CodeBuffer *code_buf, JitBloc
const GekkoOPInfo *opinfo = GetOpInfo(ops[i].inst);
js.downcountAmount += opinfo->numCycles;
if (i == (int)size - 1)
if (i == (code_block.m_num_instructions - 1))
{
js.isLastInstruction = true;
js.next_inst = 0;
@@ -708,13 +695,13 @@ const u8* JitIL::DoJit(u32 em_address, PPCAnalyst::CodeBuffer *code_buf, JitBloc
}
// Perform actual code generation
WriteCode(exitAddress);
WriteCode(nextPC);
b->codeSize = (u32)(GetCodePtr() - normalEntry);
b->originalSize = size;
b->originalSize = code_block.m_num_instructions;
#ifdef JIT_LOG_X86
LogGeneratedX86(size, code_buf, normalEntry, b);
LogGeneratedX86(code_block.m_num_instructions, code_buf, normalEntry, b);
#endif
if (SConfig::GetInstance().m_LocalCoreStartupParameter.bJITILOutputIR)
Source/Core/Core/PowerPC/JitArm32/Jit.cpp
+32 -37
View File
@@ -40,6 +40,11 @@ void JitArm::Init()
fpr.Init(this);
jo.enableBlocklink = true;
jo.optimizeGatherPipe = true;
code_block.m_stats = &js.st;
code_block.m_gpa = &js.gpa;
code_block.m_fpa = &js.fpa;
analyzer.SetOption(PPCAnalyst::PPCAnalyzer::OPTION_CONDITIONAL_CONTINUE);
}
void JitArm::ClearCache()
@@ -250,27 +255,27 @@ void JitArm::Trace()
void JitArm::PrintDebug(UGeckoInstruction inst, u32 level)
{
if (level > 0)
printf("Start: %08x OP '%s' Info\n", (u32)GetCodePtr(), PPCTables::GetInstructionName(inst));
WARN_LOG(DYNA_REC, "Start: %08x OP '%s' Info", (u32)GetCodePtr(), PPCTables::GetInstructionName(inst));
if (level > 1)
{
GekkoOPInfo* Info = GetOpInfo(inst.hex);
printf("\tOuts\n");
WARN_LOG(DYNA_REC, "\tOuts");
if (Info->flags & FL_OUT_A)
printf("\t-OUT_A: %x\n", inst.RA);
WARN_LOG(DYNA_REC, "\t-OUT_A: %x", inst.RA);
if (Info->flags & FL_OUT_D)
printf("\t-OUT_D: %x\n", inst.RD);
printf("\tIns\n");
WARN_LOG(DYNA_REC, "\t-OUT_D: %x", inst.RD);
WARN_LOG(DYNA_REC, "\tIns");
// A, AO, B, C, S
if (Info->flags & FL_IN_A)
printf("\t-IN_A: %x\n", inst.RA);
WARN_LOG(DYNA_REC, "\t-IN_A: %x", inst.RA);
if (Info->flags & FL_IN_A0)
printf("\t-IN_A0: %x\n", inst.RA);
WARN_LOG(DYNA_REC, "\t-IN_A0: %x", inst.RA);
if (Info->flags & FL_IN_B)
printf("\t-IN_B: %x\n", inst.RB);
WARN_LOG(DYNA_REC, "\t-IN_B: %x", inst.RB);
if (Info->flags & FL_IN_C)
printf("\t-IN_C: %x\n", inst.RC);
WARN_LOG(DYNA_REC, "\t-IN_C: %x", inst.RC);
if (Info->flags & FL_IN_S)
printf("\t-IN_S: %x\n", inst.RS);
WARN_LOG(DYNA_REC, "\t-IN_S: %x", inst.RS);
}
}
@@ -298,14 +303,10 @@ const u8* JitArm::DoJit(u32 em_address, PPCAnalyst::CodeBuffer *code_buf, JitBlo
// Memory exception on instruction fetch
bool memory_exception = false;
// A broken block is a block that does not end in a branch
bool broken_block = false;
if (Core::g_CoreStartupParameter.bEnableDebugging)
{
// Comment out the following to disable breakpoints (speed-up)
blockSize = 1;
broken_block = true;
Trace();
}
@@ -324,8 +325,6 @@ const u8* JitArm::DoJit(u32 em_address, PPCAnalyst::CodeBuffer *code_buf, JitBlo
}
}
int size = 0;
js.isLastInstruction = false;
js.blockStart = em_address;
js.fifoBytesThisBlock = 0;
@@ -333,17 +332,12 @@ const u8* JitArm::DoJit(u32 em_address, PPCAnalyst::CodeBuffer *code_buf, JitBlo
js.block_flags = 0;
js.cancel = false;
u32 nextPC = em_address;
// Analyze the block, collect all instructions it is made of (including inlining,
// if that is enabled), reorder instructions for optimal performance, and join joinable instructions.
u32 nextPC = em_address;
u32 merged_addresses[32];
const int capacity_of_merged_addresses = sizeof(merged_addresses) / sizeof(merged_addresses[0]);
int size_of_merged_addresses = 0;
if (!memory_exception)
{
// If there is a memory exception inside a block (broken_block==true), compile up to that instruction.
nextPC = PPCAnalyst::Flatten(em_address, &size, &js.st, &js.gpa, &js.fpa, broken_block, code_buf, blockSize, merged_addresses, capacity_of_merged_addresses, size_of_merged_addresses);
}
nextPC = analyzer.Analyze(em_address, &code_block, code_buf, blockSize);
PPCAnalyst::CodeOp *ops = code_buf->codebuffer;
const u8 *start = GetCodePtr();
@@ -398,20 +392,21 @@ const u8* JitArm::DoJit(u32 em_address, PPCAnalyst::CodeBuffer *code_buf, JitBlo
gpr.Start(js.gpa);
fpr.Start(js.fpa);
js.downcountAmount = 0;
if (!Core::g_CoreStartupParameter.bEnableDebugging)
{
for (int i = 0; i < size_of_merged_addresses; ++i)
{
const u32 address = merged_addresses[i];
js.downcountAmount += PatchEngine::GetSpeedhackCycles(address);
}
}
js.downcountAmount += PatchEngine::GetSpeedhackCycles(em_address);
js.skipnext = false;
js.blockSize = size;
js.blockSize = code_block.m_num_instructions;
js.compilerPC = nextPC;
const int DEBUG_OUTPUT = 0;
if (DEBUG_OUTPUT)
WARN_LOG(DYNA_REC, "-------0x%08x-------", em_address);
// Translate instructions
for (int i = 0; i < (int)size; i++)
for (u32 i = 0; i < code_block.m_num_instructions; i++)
{
js.compilerPC = ops[i].address;
js.op = &ops[i];
@@ -419,7 +414,7 @@ const u8* JitArm::DoJit(u32 em_address, PPCAnalyst::CodeBuffer *code_buf, JitBlo
const GekkoOPInfo *opinfo = ops[i].opinfo;
js.downcountAmount += opinfo->numCycles;
if (i == (int)size - 1)
if (i == (code_block.m_num_instructions - 1))
{
// WARNING - cmp->branch merging will screw this up.
js.isLastInstruction = true;
@@ -468,7 +463,7 @@ const u8* JitArm::DoJit(u32 em_address, PPCAnalyst::CodeBuffer *code_buf, JitBlo
}
if (!ops[i].skip)
{
PrintDebug(ops[i].inst, 0);
PrintDebug(ops[i].inst, DEBUG_OUTPUT);
if (js.memcheck && (opinfo->flags & FL_USE_FPU))
{
// Don't do this yet
@@ -485,7 +480,7 @@ const u8* JitArm::DoJit(u32 em_address, PPCAnalyst::CodeBuffer *code_buf, JitBlo
}
if (memory_exception)
BKPT(0x500);
if (broken_block)
if (code_block.m_broken)
{
printf("Broken Block going to 0x%08x\n", nextPC);
WriteExit(nextPC);
@@ -493,7 +488,7 @@ const u8* JitArm::DoJit(u32 em_address, PPCAnalyst::CodeBuffer *code_buf, JitBlo
b->flags = js.block_flags;
b->codeSize = (u32)(GetCodePtr() - normalEntry);
b->originalSize = size;
b->originalSize = code_block.m_num_instructions;
FlushIcache();
return start;
}
Source/Core/Core/PowerPC/JitArm32/Jit.h
+1
View File
@@ -42,6 +42,7 @@ private:
ArmFPRCache fpr;
PPCAnalyst::CodeBuffer code_buffer;
void DoDownCount();
void PrintDebug(UGeckoInstruction inst, u32 level);
Source/Core/Core/PowerPC/JitArm32/JitArm_Branch.cpp
+9 -16
View File
@@ -149,7 +149,6 @@ void JitArm::bcx(UGeckoInstruction inst)
INSTRUCTION_START
JITDISABLE(bJITBranchOff)
// USES_CR
_assert_msg_(DYNA_REC, js.isLastInstruction, "bcx not last instruction of block");
gpr.Flush();
fpr.Flush();
@@ -203,7 +202,8 @@ void JitArm::bcx(UGeckoInstruction inst)
if ((inst.BO & BO_DONT_DECREMENT_FLAG) == 0)
SetJumpTarget( pCTRDontBranch );
WriteExit(js.compilerPC + 4);
if (!analyzer.HasOption(PPCAnalyst::PPCAnalyzer::OPTION_CONDITIONAL_CONTINUE))
WriteExit(js.compilerPC + 4);
}
void JitArm::bcctrx(UGeckoInstruction inst)
{
@@ -265,25 +265,16 @@ void JitArm::bcctrx(UGeckoInstruction inst)
WriteExitDestInR(rA);
SetJumpTarget(b);
WriteExit(js.compilerPC + 4);
if (!analyzer.HasOption(PPCAnalyst::PPCAnalyzer::OPTION_CONDITIONAL_CONTINUE))
WriteExit(js.compilerPC + 4);
}
}
void JitArm::bclrx(UGeckoInstruction inst)
{
INSTRUCTION_START
JITDISABLE(bJITBranchOff)
if (!js.isLastInstruction &&
(inst.BO & (1 << 4)) && (inst.BO & (1 << 2))) {
if (inst.LK)
{
ARMReg rA = gpr.GetReg(false);
u32 Jumpto = js.compilerPC + 4;
MOVI2R(rA, Jumpto);
STR(rA, R9, PPCSTATE_OFF(spr[SPR_LR]));
// ARMABI_MOVI2M((u32)&LR, js.compilerPC + 4);
}
return;
}
gpr.Flush();
fpr.Flush();
@@ -342,5 +333,7 @@ void JitArm::bclrx(UGeckoInstruction inst)
SetJumpTarget( pConditionDontBranch );
if ((inst.BO & BO_DONT_DECREMENT_FLAG) == 0)
SetJumpTarget( pCTRDontBranch );
WriteExit(js.compilerPC + 4);
if (!analyzer.HasOption(PPCAnalyst::PPCAnalyzer::OPTION_CONDITIONAL_CONTINUE))
WriteExit(js.compilerPC + 4);
}
Source/Core/Core/PowerPC/JitArm32/JitArm_Integer.cpp
+3 -1
View File
@@ -1014,7 +1014,9 @@ void JitArm::twx(UGeckoInstruction inst)
SetJumpTarget(exit3);
SetJumpTarget(exit4);
SetJumpTarget(exit5);
WriteExit(js.compilerPC + 4);
if (!analyzer.HasOption(PPCAnalyst::PPCAnalyzer::OPTION_CONDITIONAL_CONTINUE))
WriteExit(js.compilerPC + 4);
gpr.Unlock(RA, RB);
}
Source/Core/Core/PowerPC/JitArmIL/JitIL.cpp
+15 -29
View File
@@ -34,6 +34,10 @@ void JitArmIL::Init()
AllocCodeSpace(CODE_SIZE);
blocks.Init();
asm_routines.Init();
code_block.m_stats = &js.st;
code_block.m_gpa = &js.gpa;
code_block.m_fpa = &js.fpa;
}
void JitArmIL::ClearCache()
@@ -202,14 +206,10 @@ const u8* JitArmIL::DoJit(u32 em_address, PPCAnalyst::CodeBuffer *code_buf, JitB
// Memory exception on instruction fetch
bool memory_exception = false;
// A broken block is a block that does not end in a branch
bool broken_block = false;
if (Core::g_CoreStartupParameter.bEnableDebugging)
{
// Comment out the following to disable breakpoints (speed-up)
blockSize = 1;
broken_block = true;
}
if (em_address == 0)
@@ -228,7 +228,6 @@ const u8* JitArmIL::DoJit(u32 em_address, PPCAnalyst::CodeBuffer *code_buf, JitB
}
int size = 0;
js.isLastInstruction = false;
js.blockStart = em_address;
js.fifoBytesThisBlock = 0;
@@ -236,17 +235,12 @@ const u8* JitArmIL::DoJit(u32 em_address, PPCAnalyst::CodeBuffer *code_buf, JitB
js.block_flags = 0;
js.cancel = false;
u32 nextPC = em_address;
// Analyze the block, collect all instructions it is made of (including inlining,
// if that is enabled), reorder instructions for optimal performance, and join joinable instructions.
u32 nextPC = em_address;
u32 merged_addresses[32];
const int capacity_of_merged_addresses = sizeof(merged_addresses) / sizeof(merged_addresses[0]);
int size_of_merged_addresses = 0;
if (!memory_exception)
{
// If there is a memory exception inside a block (broken_block==true), compile up to that instruction.
nextPC = PPCAnalyst::Flatten(em_address, &size, &js.st, &js.gpa, &js.fpa, broken_block, code_buf, blockSize, merged_addresses, capacity_of_merged_addresses, size_of_merged_addresses);
}
nextPC = analyzer.Analyze(em_address, &code_block, code_buf, blockSize);
PPCAnalyst::CodeOp *ops = code_buf->codebuffer;
const u8 *start = GetCodePtr();
@@ -271,7 +265,7 @@ const u8* JitArmIL::DoJit(u32 em_address, PPCAnalyst::CodeBuffer *code_buf, JitB
u64 codeHash = -1;
{
// For profiling and IR Writer
for (int i = 0; i < (int)size; i++)
for (u32 i = 0; i < code_block.m_num_instructions; i++)
{
const u64 inst = ops[i].inst.hex;
// Ported from boost::hash
@@ -289,19 +283,13 @@ const u8* JitArmIL::DoJit(u32 em_address, PPCAnalyst::CodeBuffer *code_buf, JitB
js.downcountAmount = 0;
if (!Core::g_CoreStartupParameter.bEnableDebugging)
{
for (int i = 0; i < size_of_merged_addresses; ++i)
{
const u32 address = merged_addresses[i];
js.downcountAmount += PatchEngine::GetSpeedhackCycles(address);
}
}
js.downcountAmount += PatchEngine::GetSpeedhackCycles(em_address);
js.skipnext = false;
js.blockSize = size;
js.blockSize = code_block.m_num_instructions;
js.compilerPC = nextPC;
// Translate instructions
for (int i = 0; i < (int)size; i++)
for (u32 i = 0; i < code_block.m_num_instructions; i++)
{
js.compilerPC = ops[i].address;
js.op = &ops[i];
@@ -309,7 +297,7 @@ const u8* JitArmIL::DoJit(u32 em_address, PPCAnalyst::CodeBuffer *code_buf, JitB
const GekkoOPInfo *opinfo = ops[i].opinfo;
js.downcountAmount += opinfo->numCycles;
if (i == (int)size - 1)
if (i == (code_block.m_num_instructions - 1))
{
// WARNING - cmp->branch merging will screw this up.
js.isLastInstruction = true;
@@ -348,21 +336,19 @@ const u8* JitArmIL::DoJit(u32 em_address, PPCAnalyst::CodeBuffer *code_buf, JitB
}
if (memory_exception)
BKPT(0x500);
if (broken_block)
if (code_block.m_broken)
{
printf("Broken Block going to 0x%08x\n", nextPC);
WriteExit(nextPC);
}
// Perform actual code generation
WriteCode(nextPC);
b->flags = js.block_flags;
b->codeSize = (u32)(GetCodePtr() - normalEntry);
b->originalSize = size;
b->originalSize = code_block.m_num_instructions;;
{
}
FlushIcache();
return start;
Source/Core/Core/PowerPC/JitCommon/JitBase.h
+3
View File
@@ -83,6 +83,9 @@ protected:
std::unordered_set<u32> fifoWriteAddresses;
};
PPCAnalyst::CodeBlock code_block;
PPCAnalyst::PPCAnalyzer analyzer;
public:
// This should probably be removed from public:
JitOptions jo;
Source/Core/Core/PowerPC/PPCAnalyst.cpp
+313 -296
View File
File diff suppressed because it is too large Load Diff
Source/Core/Core/PowerPC/PPCAnalyst.h
+83 -6
View File
@@ -64,19 +64,32 @@ struct BlockRegStats
return std::max(lastRead[reg], lastWrite[reg]) -
std::min(firstRead[reg], firstWrite[reg]);}
inline void SetInputRegister(int reg, short opindex) {
inline void SetInputRegister(int reg, short opindex)
{
if (firstRead[reg] == -1)
firstRead[reg] = (short)(opindex);
lastRead[reg] = (short)(opindex);
numReads[reg]++;
}
inline void SetOutputRegister(int reg, short opindex) {
inline void SetOutputRegister(int reg, short opindex)
{
if (firstWrite[reg] == -1)
firstWrite[reg] = (short)(opindex);
lastWrite[reg] = (short)(opindex);
numWrites[reg]++;
}
inline void Clear()
{
for (int i = 0; i < 32; ++i)
{
firstRead[i] = -1;
firstWrite[i] = -1;
numReads[i] = 0;
numWrites[i] = 0;
}
}
};
@@ -94,10 +107,74 @@ public:
};
u32 Flatten(u32 address, int *realsize, BlockStats *st, BlockRegStats *gpa,
BlockRegStats *fpa, bool &broken_block, CodeBuffer *buffer,
int blockSize, u32* merged_addresses,
int capacity_of_merged_addresses, int& size_of_merged_addresses);
struct CodeBlock
{
// Beginning PPC address.
u32 m_address;
// Number of instructions
// Gives us the size of the block.
u32 m_num_instructions;
// Some basic statistics about the block.
BlockStats *m_stats;
// Register statistics about the block.
BlockRegStats *m_gpa, *m_fpa;
// Are we a broken block?
bool m_broken;
};
class PPCAnalyzer
{
private:
void ReorderInstructions(u32 instructions, CodeOp *code);
void SetInstructionStats(CodeBlock *block, CodeOp *code, GekkoOPInfo *opinfo, u32 index);
// Options
u32 m_options;
public:
enum AnalystOption
{
// Conditional branch continuing
// If the JIT core supports conditional branches within the blocks
// Block will end on unconditional branch or other ENDBLOCK flagged instruction.
// Requires JIT support to be enabled.
OPTION_CONDITIONAL_CONTINUE = (1 << 0),
// If there is a unconditional branch that jumps to a leaf function then inline it.
// Might require JIT intervention to support it correctly.
// Requires JITBLock support for inlined code
// XXX: NOT COMPLETE
OPTION_LEAF_INLINE = (1 << 1),
// Complex blocks support jumping backwards on to themselves.
// Happens commonly in loops, pretty complex to support.
// May require register caches to use register usage metrics.
// XXX: NOT COMPLETE
OPTION_COMPLEX_BLOCK = (1 << 2),
// Similar to complex blocks.
// Instead of jumping backwards, this jumps forwards within the block.
// Requires JIT support to work.
// XXX: NOT COMPLETE
OPTION_FORWARD_JUMP = (1 << 3),
};
PPCAnalyzer() : m_options(0) {}
// Option setting/getting
void SetOption(AnalystOption option) { m_options |= option; }
void ClearOption(AnalystOption option) { m_options &= ~(option); }
bool HasOption(AnalystOption option) { return !!(m_options & option); }
u32 Analyze(u32 address, CodeBlock *block, CodeBuffer *buffer, u32 blockSize);
};
void LogFunctionCall(u32 addr);
void FindFunctions(u32 startAddr, u32 endAddr, PPCSymbolDB *func_db);
bool AnalyzeFunction(u32 startAddr, Symbol &func, int max_size = 0);
Source/Core/DolphinWX/Debugger/JitWindow.cpp
+12 -10
View File
@@ -129,8 +129,7 @@ void CJitWindow::Compare(u32 em_address)
const u8 *code = (const u8 *)jit->GetBlockCache()->GetCompiledCodeFromBlock(block_num);
u64 disasmPtr = (u64)code;
int size = block->codeSize;
const u8 *end = code + size;
const u8 *end = code + block->codeSize;
char *sptr = (char*)xDis;
int num_x86_instructions = 0;
@@ -154,14 +153,17 @@ void CJitWindow::Compare(u32 em_address)
PPCAnalyst::BlockStats st;
PPCAnalyst::BlockRegStats gpa;
PPCAnalyst::BlockRegStats fpa;
bool broken_block = false;
u32 merged_addresses[32];
const int capacity_of_merged_addresses = sizeof(merged_addresses) / sizeof(merged_addresses[0]);
int size_of_merged_addresses;
if (PPCAnalyst::Flatten(ppc_addr, &size, &st, &gpa, &fpa, broken_block, &code_buffer, size, merged_addresses, capacity_of_merged_addresses, size_of_merged_addresses) != 0xffffffff)
PPCAnalyst::CodeBlock code_block;
PPCAnalyst::PPCAnalyzer analyzer;
code_block.m_stats = &st;
code_block.m_gpa = &gpa;
code_block.m_fpa = &fpa;
if (analyzer.Analyze(ppc_addr, &code_block, &code_buffer, block->codeSize) != 0xFFFFFFFF)
{
sptr = (char*)xDis;
for (int i = 0; i < size; i++)
for (u32 i = 0; i < code_block.m_num_instructions; i++)
{
const PPCAnalyst::CodeOp &op = code_buffer.codebuffer[i];
char temp[256];
@@ -181,9 +183,9 @@ void CJitWindow::Compare(u32 em_address)
sptr += sprintf(sptr, "%i estimated cycles\n", st.numCycles);
sptr += sprintf(sptr, "Num instr: PPC: %i x86: %i (blowup: %i%%)\n",
size, num_x86_instructions, 100 * (num_x86_instructions / size - 1));
code_block.m_num_instructions, num_x86_instructions, 100 * (num_x86_instructions / code_block.m_num_instructions - 1));
sptr += sprintf(sptr, "Num bytes: PPC: %i x86: %i (blowup: %i%%)\n",
size * 4, block->codeSize, 100 * (block->codeSize / (4 * size) - 1));
code_block.m_num_instructions * 4, block->codeSize, 100 * (block->codeSize / (4 * code_block.m_num_instructions) - 1));
ppc_box->SetValue(StrToWxStr((char*)xDis));
}