ShuriZma
/
dolphin
mirror of https://github.com/dolphin-emu/dolphin.git
1
0
Fork 0
Code Issues Packages Projects Releases Wiki Activity

Merge pull request #10046 from JosJuice/jitarm64-improve-const-stores 

JitArm64: FIFO optimization improvements
This commit is contained in:
JMC47 2022-10-22 05:03:43 -04:00 committed by GitHub
parent bc4d08047d 351d095fff
commit 2153276ab9
No known key found for this signature in database
GPG Key ID: 4AEE18F83AFDEB23
4 changed files with 92 additions and 31 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

58
Source/Core/Core/PowerPC/JitArm64/Jit.cpp
View File

@ -320,6 +320,51 @@ void JitArm64::FreeStack()
#endif
}
void JitArm64::IntializeSpeculativeConstants()
{
// If the block depends on an input register which looks like a gather pipe or MMIO related
// constant, guess that it is actually a constant input, and specialize the block based on this
// assumption. This happens when there are branches in code writing to the gather pipe, but only
// the first block loads the constant.
// Insert a check at the start of the block to verify that the value is actually constant.
// This can save a lot of backpatching and optimize gather pipe writes in more places.
const u8* fail = nullptr;
for (auto i : code_block.m_gpr_inputs)
{
u32 compile_time_value = PowerPC::ppcState.gpr[i];
if (PowerPC::IsOptimizableGatherPipeWrite(compile_time_value) ||
PowerPC::IsOptimizableGatherPipeWrite(compile_time_value - 0x8000) ||
compile_time_value == 0xCC000000)
{
if (!fail)
{
SwitchToFarCode();
fail = GetCodePtr();
MOVI2R(DISPATCHER_PC, js.blockStart);
STR(IndexType::Unsigned, DISPATCHER_PC, PPC_REG, PPCSTATE_OFF(pc));
MOVP2R(ARM64Reg::X8, &JitInterface::CompileExceptionCheck);
MOVI2R(ARM64Reg::W0, static_cast<u32>(JitInterface::ExceptionType::SpeculativeConstants));
// Write dispatcher_no_check to LR for tail call
MOVP2R(ARM64Reg::X30, dispatcher_no_check);
BR(ARM64Reg::X8);
SwitchToNearCode();
}
ARM64Reg tmp = gpr.GetReg();
ARM64Reg value = gpr.R(i);
MOVI2R(tmp, compile_time_value);
CMP(value, tmp);
gpr.Unlock(tmp);
FixupBranch no_fail = B(CCFlags::CC_EQ);
B(fail);
SetJumpTarget(no_fail);
gpr.SetImmediate(i, compile_time_value, true);
}
}
}
void JitArm64::WriteExit(u32 destination, bool LK, u32 exit_address_after_return)
{
Cleanup();
@ -793,10 +838,11 @@ bool JitArm64::DoJit(u32 em_address, JitBlock* b, u32 nextPC)
SetJumpTarget(fail);
MOVI2R(DISPATCHER_PC, js.blockStart);
STR(IndexType::Unsigned, DISPATCHER_PC, PPC_REG, PPCSTATE_OFF(pc));
MOVP2R(ARM64Reg::X8, &JitInterface::CompileExceptionCheck);
MOVI2R(ARM64Reg::W0, static_cast<u32>(JitInterface::ExceptionType::PairedQuantize));
MOVP2R(ARM64Reg::X1, &JitInterface::CompileExceptionCheck);
BLR(ARM64Reg::X1);
B(dispatcher_no_check);
// Write dispatcher_no_check to LR for tail call
MOVP2R(ARM64Reg::X30, dispatcher_no_check);
BR(ARM64Reg::X8);
SwitchToNearCode();
SetJumpTarget(no_fail);
js.assumeNoPairedQuantize = true;
@ -806,6 +852,12 @@ bool JitArm64::DoJit(u32 em_address, JitBlock* b, u32 nextPC)
gpr.Start(js.gpa);
fpr.Start(js.fpa);
if (js.noSpeculativeConstantsAddresses.find(js.blockStart) ==
js.noSpeculativeConstantsAddresses.end())
{
IntializeSpeculativeConstants();
}
// Translate instructions
for (u32 i = 0; i < code_block.m_num_instructions; i++)
{

2
Source/Core/Core/PowerPC/JitArm64/Jit.h
View File

@ -285,6 +285,8 @@ protected:
void ResetFreeMemoryRanges();
void IntializeSpeculativeConstants();
// AsmRoutines
void GenerateAsm();
void GenerateCommonAsm();

56
Source/Core/Core/PowerPC/JitArm64/JitArm64_RegCache.cpp
View File

@ -202,37 +202,40 @@ void Arm64GPRCache::FlushRegister(size_t index, bool maintain_state, ARM64Reg tm
}
else if (reg.GetType() == RegType::Immediate)
{
if (!reg.GetImm())
if (reg.IsDirty())
{
m_emit->STR(IndexType::Unsigned, bitsize == 64 ? ARM64Reg::ZR : ARM64Reg::WZR, PPC_REG,
u32(guest_reg.ppc_offset));
}
else
{
bool allocated_tmp_reg = false;
if (tmp_reg != ARM64Reg::INVALID_REG)
if (!reg.GetImm())
{
ASSERT(IsGPR(tmp_reg));
m_emit->STR(IndexType::Unsigned, bitsize == 64 ? ARM64Reg::ZR : ARM64Reg::WZR, PPC_REG,
u32(guest_reg.ppc_offset));
}
else
{
ASSERT_MSG(DYNA_REC, !maintain_state,
"Flushing immediate while maintaining state requires temporary register");
tmp_reg = GetReg();
allocated_tmp_reg = true;
bool allocated_tmp_reg = false;
if (tmp_reg != ARM64Reg::INVALID_REG)
{
ASSERT(IsGPR(tmp_reg));
}
else
{
ASSERT_MSG(DYNA_REC, !maintain_state,
"Flushing immediate while maintaining state requires temporary register");
tmp_reg = GetReg();
allocated_tmp_reg = true;
}
const ARM64Reg encoded_tmp_reg = bitsize != 64 ? tmp_reg : EncodeRegTo64(tmp_reg);
m_emit->MOVI2R(encoded_tmp_reg, reg.GetImm());
m_emit->STR(IndexType::Unsigned, encoded_tmp_reg, PPC_REG, u32(guest_reg.ppc_offset));
if (allocated_tmp_reg)
UnlockRegister(tmp_reg);
}
const ARM64Reg encoded_tmp_reg = bitsize != 64 ? tmp_reg : EncodeRegTo64(tmp_reg);
m_emit->MOVI2R(encoded_tmp_reg, reg.GetImm());
m_emit->STR(IndexType::Unsigned, encoded_tmp_reg, PPC_REG, u32(guest_reg.ppc_offset));
if (allocated_tmp_reg)
UnlockRegister(tmp_reg);
if (!maintain_state)
reg.Flush();
}
if (!maintain_state)
reg.Flush();
}
}
@ -335,12 +338,13 @@ ARM64Reg Arm64GPRCache::R(const GuestRegInfo& guest_reg)
return ARM64Reg::INVALID_REG;
}
void Arm64GPRCache::SetImmediate(const GuestRegInfo& guest_reg, u32 imm)
void Arm64GPRCache::SetImmediate(const GuestRegInfo& guest_reg, u32 imm, bool dirty)
{
OpArg& reg = guest_reg.reg;
if (reg.GetType() == RegType::Register)
UnlockRegister(EncodeRegTo32(reg.GetReg()));
reg.LoadToImm(imm);
reg.SetDirty(dirty);
}
void Arm64GPRCache::BindToRegister(const GuestRegInfo& guest_reg, bool will_read, bool will_write)
@ -373,8 +377,8 @@ void Arm64GPRCache::BindToRegister(const GuestRegInfo& guest_reg, bool will_read
m_emit->MOVI2R(host_reg, reg.GetImm());
}
reg.Load(host_reg);
// If the register had an immediate value, the register was effectively already dirty
reg.SetDirty(true);
if (will_write)
reg.SetDirty(true);
}
else if (will_write)
{

7
Source/Core/Core/PowerPC/JitArm64/JitArm64_RegCache.h
View File

@ -262,7 +262,10 @@ public:
Arm64Gen::ARM64Reg CR(size_t preg) { return R(GetGuestCR(preg)); }
// Set a register to an immediate. Only valid for guest GPRs.
void SetImmediate(size_t preg, u32 imm) { SetImmediate(GetGuestGPR(preg), imm); }
void SetImmediate(size_t preg, u32 imm, bool dirty = true)
{
SetImmediate(GetGuestGPR(preg), imm, dirty);
}
// Returns if a register is set as an immediate. Only valid for guest GPRs.
bool IsImm(size_t preg) const { return GetGuestGPROpArg(preg).GetType() == RegType::Immediate; }
@ -345,7 +348,7 @@ private:
GuestRegInfo GetGuestByIndex(size_t index);
Arm64Gen::ARM64Reg R(const GuestRegInfo& guest_reg);
void SetImmediate(const GuestRegInfo& guest_reg, u32 imm);
void SetImmediate(const GuestRegInfo& guest_reg, u32 imm, bool dirty);
void BindToRegister(const GuestRegInfo& guest_reg, bool will_read, bool will_write = true);
void FlushRegisters(BitSet32 regs, bool maintain_state, Arm64Gen::ARM64Reg tmp_reg);