The-Homebrew-Cloud/dolphin
2
0
Fork 0
mirror of https://github.com/dolphin-emu/dolphin.git synced 2025-01-09 15:49:25 +01:00
Code Issues Packages Projects Releases Wiki Activity

JitArm64: Get rid of one MOV from lmw/stmw

Browse Source 
This is possible now that EmitBackpatchRoutine lets the caller choose
addr register.

Note: The removed MOV is actually traded for one added MOV in farcode.
This commit is contained in:
JosJuice 2024-12-29 21:28:59 +01:00
parent c3b6e67fc7
commit c88c6f5f18
1 changed files with 66 additions and 31 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

97
Source/Core/Core/PowerPC/JitArm64/JitArm64_LoadStore.cpp
View File

@ -519,21 +519,26 @@ void JitArm64::lmw(UGeckoInstruction inst)
// MMU games make use of a >= d despite this being invalid according to the PEM. // MMU games make use of a >= d despite this being invalid according to the PEM.
// If a >= d occurs, we must make sure to not re-read rA after starting doing the loads. // If a >= d occurs, we must make sure to not re-read rA after starting doing the loads.
ARM64Reg addr_reg = ARM64Reg::W1; ARM64Reg addr_reg = ARM64Reg::W1;
Arm64RegCache::ScopedARM64Reg addr_base_reg;
bool a_is_addr_base_reg = false; bool a_is_addr_base_reg = false;
if (!a) if (!a)
MOVI2R(addr_reg, offset);
else if (gpr.IsImm(a))
MOVI2R(addr_reg, gpr.GetImm(a) + offset);
else if (a < d && offset + (31 - d) * 4 < 0x1000)
a_is_addr_base_reg = true;
else
ADDI2R(addr_reg, gpr.R(a), offset, addr_reg);
Arm64RegCache::ScopedARM64Reg addr_base_reg;
if (!a_is_addr_base_reg)
{ {
addr_base_reg = gpr.GetScopedReg(); addr_base_reg = gpr.GetScopedReg();
MOV(addr_base_reg, addr_reg); MOVI2R(addr_base_reg, offset);
}
else if (gpr.IsImm(a))
{
addr_base_reg = gpr.GetScopedReg();
MOVI2R(addr_base_reg, gpr.GetImm(a) + offset);
}
else if (a < d && offset + (31 - d) * 4 < 0x1000)
{
a_is_addr_base_reg = true;
}
else
{
addr_base_reg = gpr.GetScopedReg();
ADDI2R(addr_base_reg, gpr.R(a), offset, addr_base_reg);
} }
BitSet32 gprs_to_discard{}; BitSet32 gprs_to_discard{};
@ -576,11 +581,23 @@ void JitArm64::lmw(UGeckoInstruction inst)
{ {
gpr.BindToRegister(i, false, false); gpr.BindToRegister(i, false, false);
ARM64Reg dest_reg = gpr.R(i); ARM64Reg dest_reg = gpr.R(i);
ARM64Reg current_iteration_addr_reg = addr_reg;
if (a_is_addr_base_reg) if (a_is_addr_base_reg)
ADDI2R(addr_reg, gpr.R(a), offset + (i - d) * 4); {
else if (i != d) const u32 current_iteration_offset = offset + (i - d) * 4;
ADDI2R(addr_reg, addr_base_reg, (i - d) * 4); if (current_iteration_offset != 0)
ADDI2R(addr_reg, gpr.R(a), current_iteration_offset);
else
current_iteration_addr_reg = gpr.R(a);
}
else
{
if (i != d)
ADDI2R(addr_reg, addr_base_reg, (i - d) * 4);
else
current_iteration_addr_reg = addr_base_reg;
}
BitSet32 scratch_gprs; BitSet32 scratch_gprs;
BitSet32 scratch_fprs; BitSet32 scratch_fprs;
@ -588,8 +605,8 @@ void JitArm64::lmw(UGeckoInstruction inst)
if (jo.memcheck) if (jo.memcheck)
scratch_gprs[DecodeReg(ARM64Reg::W0)] = true; scratch_gprs[DecodeReg(ARM64Reg::W0)] = true;
EmitBackpatchRoutine(flags, MemAccessMode::Auto, dest_reg, addr_reg, scratch_gprs, EmitBackpatchRoutine(flags, MemAccessMode::Auto, dest_reg, current_iteration_addr_reg,
scratch_fprs); scratch_gprs, scratch_fprs);
gpr.BindToRegister(i, false, true); gpr.BindToRegister(i, false, true);
ASSERT(dest_reg == gpr.R(i)); ASSERT(dest_reg == gpr.R(i));
@ -633,21 +650,26 @@ void JitArm64::stmw(UGeckoInstruction inst)
gpr.Lock(ARM64Reg::W2, ARM64Reg::W30); gpr.Lock(ARM64Reg::W2, ARM64Reg::W30);
ARM64Reg addr_reg = ARM64Reg::W2; ARM64Reg addr_reg = ARM64Reg::W2;
Arm64RegCache::ScopedARM64Reg addr_base_reg;
bool a_is_addr_base_reg = false; bool a_is_addr_base_reg = false;
if (!a) if (!a)
MOVI2R(addr_reg, offset);
else if (gpr.IsImm(a))
MOVI2R(addr_reg, gpr.GetImm(a) + offset);
else if (offset + (31 - s) * 4 < 0x1000)
a_is_addr_base_reg = true;
else
ADDI2R(addr_reg, gpr.R(a), offset, addr_reg);
Arm64GPRCache::ScopedARM64Reg addr_base_reg;
if (!a_is_addr_base_reg)
{ {
addr_base_reg = gpr.GetScopedReg(); addr_base_reg = gpr.GetScopedReg();
MOV(addr_base_reg, addr_reg); MOVI2R(addr_base_reg, offset);
}
else if (gpr.IsImm(a))
{
addr_base_reg = gpr.GetScopedReg();
MOVI2R(addr_base_reg, gpr.GetImm(a) + offset);
}
else if (offset + (31 - s) * 4 < 0x1000)
{
a_is_addr_base_reg = true;
}
else
{
addr_base_reg = gpr.GetScopedReg();
ADDI2R(addr_base_reg, gpr.R(a), offset, addr_base_reg);
} }
BitSet32 gprs_to_discard{}; BitSet32 gprs_to_discard{};
@ -690,17 +712,30 @@ void JitArm64::stmw(UGeckoInstruction inst)
for (u32 i = s; i < 32; i++) for (u32 i = s; i < 32; i++)
{ {
ARM64Reg src_reg = gpr.R(i); ARM64Reg src_reg = gpr.R(i);
ARM64Reg current_iteration_addr_reg = addr_reg;
if (a_is_addr_base_reg) if (a_is_addr_base_reg)
ADDI2R(addr_reg, gpr.R(a), offset + (i - s) * 4); {
else if (i != s) const u32 current_iteration_offset = offset + (i - s) * 4;
ADDI2R(addr_reg, addr_base_reg, (i - s) * 4); if (current_iteration_offset != 0)
ADDI2R(addr_reg, gpr.R(a), current_iteration_offset);
else
current_iteration_addr_reg = gpr.R(a);
}
else
{
if (i != s)
ADDI2R(addr_reg, addr_base_reg, (i - s) * 4);
else
current_iteration_addr_reg = addr_base_reg;
}
BitSet32 scratch_gprs; BitSet32 scratch_gprs;
BitSet32 scratch_fprs; BitSet32 scratch_fprs;
scratch_gprs[DecodeReg(addr_reg)] = true; scratch_gprs[DecodeReg(addr_reg)] = true;
EmitBackpatchRoutine(flags, MemAccessMode::Auto, src_reg, addr_reg, scratch_gprs, scratch_fprs); EmitBackpatchRoutine(flags, MemAccessMode::Auto, src_reg, current_iteration_addr_reg,
scratch_gprs, scratch_fprs);
// To reduce register pressure and to avoid getting a pipeline-unfriendly long run of stores // To reduce register pressure and to avoid getting a pipeline-unfriendly long run of stores
// after this instruction, flush registers that would be flushed after this instruction anyway. // after this instruction, flush registers that would be flushed after this instruction anyway.