JitArm64: Optimize dcbx being called in a loop over a large memory region

JitArm64 port of 8b2f5d5.
This commit is contained in:
JosJuice 2021-08-13 14:21:27 +02:00
parent d162015112
commit 856ff296ae
1 changed files with 99 additions and 8 deletions

View File

@ -553,21 +553,97 @@ void JitArm64::dcbx(UGeckoInstruction inst)
INSTRUCTION_START
JITDISABLE(bJITLoadStoreOff);
u32 a = inst.RA, b = inst.RB;
// Check if the next instructions match a known looping pattern:
// - dcbx rX
// - addi rX,rX,32
// - bdnz+ -8
const bool make_loop = a == 0 && b != 0 && CanMergeNextInstructions(2) &&
(js.op[1].inst.hex & 0xfc00'ffff) == 0x38000020 &&
js.op[1].inst.RA_6 == b && js.op[1].inst.RD_2 == b &&
js.op[2].inst.hex == 0x4200fff8;
gpr.Lock(ARM64Reg::W0, ARM64Reg::W30);
if (make_loop)
gpr.Lock(ARM64Reg::W1);
ARM64Reg WA = ARM64Reg::W30;
if (make_loop)
gpr.BindToRegister(b, true);
ARM64Reg loop_counter = ARM64Reg::INVALID_REG;
if (make_loop)
{
// We'll execute somewhere between one single cacheline invalidation and however many are needed
// to reduce the downcount to zero, never exceeding the amount requested by the game.
// To stay consistent with the rest of the code we adjust the involved registers (CTR and Rb)
// by the amount of cache lines we invalidate minus one -- since we'll run the regular addi and
// bdnz afterwards! So if we invalidate a single cache line, we don't adjust the registers at
// all, if we invalidate 2 cachelines we adjust the registers by one step, and so on.
ARM64Reg reg_cycle_count = gpr.GetReg();
ARM64Reg reg_downcount = gpr.GetReg();
loop_counter = ARM64Reg::W1;
ARM64Reg WB = ARM64Reg::W0;
// Figure out how many loops we want to do.
const u8 cycle_count_per_loop =
js.op[0].opinfo->numCycles + js.op[1].opinfo->numCycles + js.op[2].opinfo->numCycles;
LDR(IndexType::Unsigned, reg_downcount, PPC_REG, PPCSTATE_OFF(downcount));
MOVI2R(WA, 0);
CMP(reg_downcount, 0); // if (downcount <= 0)
FixupBranch downcount_is_zero_or_negative = B(CCFlags::CC_LE); // only do 1 invalidation; else:
LDR(IndexType::Unsigned, loop_counter, PPC_REG, PPCSTATE_OFF_SPR(SPR_CTR));
MOVI2R(reg_cycle_count, cycle_count_per_loop);
SDIV(WB, reg_downcount, reg_cycle_count); // WB = downcount / cycle_count
SUB(WA, loop_counter, 1); // WA = CTR - 1
// ^ Note that this CTR-1 implicitly handles the CTR == 0 case correctly.
CMP(WB, WA);
CSEL(WA, WB, WA, CCFlags::CC_LO); // WA = min(WB, WA)
// WA now holds the amount of loops to execute minus 1, which is the amount we need to adjust
// downcount, CTR, and Rb by to exit the loop construct with the right values in those
// registers.
// CTR -= WA
SUB(loop_counter, loop_counter, WA);
STR(IndexType::Unsigned, loop_counter, PPC_REG, PPCSTATE_OFF_SPR(SPR_CTR));
// downcount -= (WA * reg_cycle_count)
MUL(WB, WA, reg_cycle_count);
// ^ Note that this cannot overflow because it's limited by (downcount/cycle_count).
SUB(reg_downcount, reg_downcount, WB);
STR(IndexType::Unsigned, reg_downcount, PPC_REG, PPCSTATE_OFF(downcount));
SetJumpTarget(downcount_is_zero_or_negative);
// Load the loop_counter register with the amount of invalidations to execute.
ADD(loop_counter, WA, 1);
gpr.Unlock(reg_cycle_count, reg_downcount);
}
ARM64Reg effective_addr = ARM64Reg::W0;
ARM64Reg physical_addr = MSR.IR ? gpr.GetReg() : effective_addr;
ARM64Reg value = gpr.GetReg();
ARM64Reg WA = ARM64Reg::W30;
u32 a = inst.RA, b = inst.RB;
if (a)
ADD(effective_addr, gpr.R(a), gpr.R(b));
else
MOV(effective_addr, gpr.R(b));
if (make_loop)
{
// This is the best place to adjust Rb to what it should be since WA still has the
// adjusted loop count and we're done reading from Rb.
ADD(gpr.R(b), gpr.R(b), WA, ArithOption(WA, ShiftType::LSL, 5)); // Rb += (WA * 32)
}
// Translate effective address to physical address.
const u8* loop_start = GetCodePtr();
FixupBranch bat_lookup_failed;
if (MSR.IR)
{
@ -586,9 +662,18 @@ void JitArm64::dcbx(UGeckoInstruction inst)
LSRV(value, value, WA); // move current bit to bit 0
FixupBranch bit_not_set = TBZ(value, 0);
FixupBranch far_addr = B();
FixupBranch invalidate_needed = B();
SetJumpTarget(bit_not_set);
if (make_loop)
{
ADD(effective_addr, effective_addr, 32);
SUBS(loop_counter, loop_counter, 1);
B(CCFlags::CC_NEQ, loop_start);
}
SwitchToFarCode();
SetJumpTarget(far_addr);
SetJumpTarget(invalidate_needed);
if (MSR.IR)
SetJumpTarget(bat_lookup_failed);
@ -598,12 +683,17 @@ void JitArm64::dcbx(UGeckoInstruction inst)
gprs_to_push[DecodeReg(physical_addr)] = false;
gprs_to_push[DecodeReg(value)] = false;
gprs_to_push[DecodeReg(WA)] = false;
if (make_loop)
gprs_to_push[DecodeReg(loop_counter)] = false;
ABI_PushRegisters(gprs_to_push);
m_float_emit.ABI_PushRegisters(fprs_to_push, ARM64Reg::X30);
// W0 (the function call argument) was already set earlier
MOVP2R(ARM64Reg::X8, &JitInterface::InvalidateICacheLine);
// The function call arguments are already in the correct registers
if (make_loop)
MOVP2R(ARM64Reg::X8, &JitInterface::InvalidateICacheLines);
else
MOVP2R(ARM64Reg::X8, &JitInterface::InvalidateICacheLine);
BLR(ARM64Reg::X8);
m_float_emit.ABI_PopRegisters(fprs_to_push, ARM64Reg::X30);
@ -611,12 +701,13 @@ void JitArm64::dcbx(UGeckoInstruction inst)
FixupBranch near_addr = B();
SwitchToNearCode();
SetJumpTarget(bit_not_set);
SetJumpTarget(near_addr);
gpr.Unlock(effective_addr, value, WA);
if (MSR.IR)
gpr.Unlock(physical_addr);
if (make_loop)
gpr.Unlock(loop_counter);
}
void JitArm64::dcbt(UGeckoInstruction inst)