Document register allocator and return stack buffer optimization
This commit is contained in:
MerryMage
parent
95f34c683c
commit
ff00b8c555
2 changed files with 232 additions and 0 deletions
85
docs/RegisterAllocator.md
Normal file
85
docs/RegisterAllocator.md
Normal file
|
|
@ -0,0 +1,85 @@
|
||||||
|
# Register Allocation (x64 Backend)
|
||||||
|
|
||||||
|
`HostLoc`s contain values. A `HostLoc` ("host value location") is either a host CPU register or a host spill location.
|
||||||
|
|
||||||
|
Values once set cannot be changed. Values can however be moved by the register allocator between `HostLoc`s. This is handled by the register allocator itself and code that uses the register allocator need not and should not move values between registers.
|
||||||
|
|
||||||
|
The register allocator is based on three concepts: `Use`, `Def` and `Scratch`.
|
||||||
|
|
||||||
|
* `Use`: The use of a value.
|
||||||
|
* `Def`: The definition of a value, this is the only time when a value is set.
|
||||||
|
* `Scratch`: Allocate a register that can be freely modified as one wishes.
|
||||||
|
|
||||||
|
Note that `Use`ing a value decrements its `use_count` by one. When the `use_count` reaches zero the value is discarded and no longer exists.
|
||||||
|
|
||||||
|
The member functions on `RegAlloc` are just a combination of the above concepts.
|
||||||
|
|
||||||
|
### `Scratch`
|
||||||
|
|
||||||
|
Xbyak::Reg64 ScratchGpr(HostLocList desired_locations = any_gpr)
|
||||||
|
Xbyak::Xmm ScratchXmm(HostLocList desired_locations = any_xmm)
|
||||||
|
|
||||||
|
At runtime, allocate one of the registers in `desired_locations`. You are free to modify the register. The register is discarded at the end of the allocation scope.
|
||||||
|
|
||||||
|
### Pure `Use`
|
||||||
|
|
||||||
|
Xbyak::Reg64 UseGpr(IR::Value use_value, HostLocList desired_locations = any_gpr);
|
||||||
|
Xbyak::Xmm UseXmm(IR::Value use_value, HostLocList desired_locations = any_xmm);
|
||||||
|
OpArg UseOpArg(IR::Value use_value, HostLocList desired_locations);
|
||||||
|
|
||||||
|
At runtime, the value corresponding to `use_value` will be placed into one of the `HostLoc`s specified by `desired_locations`. The return value is the actual location.
|
||||||
|
|
||||||
|
This register **must not** have it's value changed.
|
||||||
|
|
||||||
|
* `UseGpr`: The location is a GPR.
|
||||||
|
* `UseXmm`: The location is an XMM register.
|
||||||
|
* `UseOpArg`: The location may be one of the locations specified by `desired_locations`, but may also be a host memory reference.
|
||||||
|
|
||||||
|
### `UseScratch`
|
||||||
|
|
||||||
|
Xbyak::Reg64 UseScratchGpr(IR::Value use_value, HostLocList desired_locations = any_gpr)
|
||||||
|
Xbyak::Xmm UseScratchXmm(IR::Value use_value, HostLocList desired_locations = any_xmm)
|
||||||
|
|
||||||
|
At runtime, the value corresponding to `use_value` will be placed into one of the `HostLoc`s specified by `desired_locations`. The return value is the actual location.
|
||||||
|
|
||||||
|
You are free to modify the register. The register is discarded at the end of the allocation scope.
|
||||||
|
|
||||||
|
### `Def`
|
||||||
|
|
||||||
|
A `Def` is the defintion of a value. This is the only time when a value may be set.
|
||||||
|
|
||||||
|
Xbyak::Xmm DefXmm(IR::Inst* def_inst, HostLocList desired_locations = any_xmm)
|
||||||
|
Xbyak::Reg64 DefGpr(IR::Inst* def_inst, HostLocList desired_locations = any_gpr)
|
||||||
|
|
||||||
|
By calling `DefXmm` or `DefGpr`, you are stating that you wish to define the value for `def_inst`, and you wish to write the value to one of the `HostLoc`s specified by `desired_locations`. You must write the value to the register returned.
|
||||||
|
|
||||||
|
### `AddDef`
|
||||||
|
|
||||||
|
Adding a `Def` to an existing value.
|
||||||
|
|
||||||
|
void RegisterAddDef(IR::Inst* def_inst, const IR::Value& use_inst);
|
||||||
|
|
||||||
|
You are declaring that the value for `def_inst` is the same as the value for `use_inst`. No host machine instructions are emitted.
|
||||||
|
|
||||||
|
### `UseDef`
|
||||||
|
|
||||||
|
Xbyak::Reg64 UseDefGpr(IR::Value use_value, IR::Inst* def_inst, HostLocList desired_locations = any_gpr)
|
||||||
|
Xbyak::Xmm UseDefXmm(IR::Value use_value, IR::Inst* def_inst, HostLocList desired_locations = any_xmm)
|
||||||
|
|
||||||
|
At runtime, the value corresponding to `use_value` will be placed into one of the `HostLoc`s specified by `desired_locations`. The return value is the actual location. You must write the value correponding to `def_inst` by the end of the allocation scope.
|
||||||
|
|
||||||
|
### `UseDef` (OpArg variant)
|
||||||
|
|
||||||
|
std::tuple<OpArg, Xbyak::Reg64> UseDefOpArgGpr(IR::Value use_value, IR::Inst* def_inst, HostLocList desired_locations = any_gpr)
|
||||||
|
std::tuple<OpArg, Xbyak::Xmm> UseDefOpArgXmm(IR::Value use_value, IR::Inst* def_inst, HostLocList desired_locations = any_xmm)
|
||||||
|
|
||||||
|
These have the same semantics as `UseDefGpr` and `UseDefXmm` except `use_value` may not be present in the register, and may actually be in a host memory location.
|
||||||
|
|
||||||
|
## When to use each?
|
||||||
|
|
||||||
|
The variety of different ways to `Use` and `Def` values are for performance reasons.
|
||||||
|
|
||||||
|
* `UseDef`: Instead of performing a `Use` and a `Def`, `UseDef` uses one less register in the case when this `Use` is the last `Use` of a value.
|
||||||
|
* `UseScratch`: Instead of performing a `Use` and a `Scratch`, `UseScratch` uses one less register in the case when this `Use` is the last `Use` of a value.
|
||||||
|
* `AddDef`: This drastically reduces the number of registers required when it can be used. It can be used when values are truncations of other values. For example, if `u8_value` contains the truncation of `u32_value`, `AddDef(u8_value, u32_value)` is a valid definition of `u8_value`.
|
||||||
|
* OpArg variants: Save host code-cache by merging memory loads into other instructions instead of the register allocator having to emit a `mov`.
|
||||||
147
docs/ReturnStackBufferOptimization.md
Normal file
147
docs/ReturnStackBufferOptimization.md
Normal file
|
|
@ -0,0 +1,147 @@
|
||||||
|
# Return Stack Buffer Optimization (x64 Backend)
|
||||||
|
|
||||||
|
One of the optimizations that dynarmic does is block-linking. Block-linking is done when
|
||||||
|
the destination address of a jump is available at JIT-time. Instead of returning to the
|
||||||
|
dispatcher at the end of a block we can perform block-linking: just jump directly to the
|
||||||
|
next block. This is beneficial because returning to the dispatcher can often be quite
|
||||||
|
expensive.
|
||||||
|
|
||||||
|
What should we do in cases when we can't predict the destination address? The eponymous
|
||||||
|
example is when executing a return statement at the end of a function; the return address
|
||||||
|
is not statically known at compile time.
|
||||||
|
|
||||||
|
We deal with this by using a return stack buffer: When we execute a call instruction,
|
||||||
|
we push our prediction onto the RSB. When we execute a return instruction, we pop a
|
||||||
|
prediction off the RSB. If the prediction is a hit, we immediately jump to the relevant
|
||||||
|
compiled block. Otherwise, we return to the dispatcher.
|
||||||
|
|
||||||
|
This is the essential idea behind this optimization.
|
||||||
|
|
||||||
|
## `UniqueHash`
|
||||||
|
|
||||||
|
One complication dynarmic has is that a compiled block is not uniquely identifiable by
|
||||||
|
the PC alone, but bits in the FPSCR and CPSR are also relevant. We resolve this by
|
||||||
|
computing a 64-bit `UniqueHash` that is guaranteed to uniquely identify a block.
|
||||||
|
|
||||||
|
u64 LocationDescriptor::UniqueHash() const {
|
||||||
|
// This value MUST BE UNIQUE.
|
||||||
|
// This calculation has to match up with EmitX64::EmitTerminalPopRSBHint
|
||||||
|
u64 pc_u64 = u64(arm_pc);
|
||||||
|
u64 fpscr_u64 = u64(fpscr.Value()) << 32;
|
||||||
|
u64 t_u64 = cpsr.T() ? (1ull << 35) : 0;
|
||||||
|
u64 e_u64 = cpsr.E() ? (1ull << 39) : 0;
|
||||||
|
return pc_u64 | fpscr_u64 | t_u64 | e_u64;
|
||||||
|
}
|
||||||
|
|
||||||
|
## Our implementation isn't actually a stack
|
||||||
|
|
||||||
|
Dynarmic's RSB isn't actually a stack. It was implemented as a ring buffer because
|
||||||
|
that showed better performance in tests.
|
||||||
|
|
||||||
|
### RSB Structure
|
||||||
|
|
||||||
|
The RSB is implemented as a ring buffer. `rsb_ptr` is the index of the insertion
|
||||||
|
point. Each element in `rsb_location_descriptors` is a `UniqueHash` and they
|
||||||
|
each correspond to an element in `rsb_codeptrs`. `rsb_codeptrs` contains the
|
||||||
|
host addresses for the corresponding the compiled blocks.
|
||||||
|
|
||||||
|
`RSBSize` was chosen by performance testing. Note that this is bigger than the
|
||||||
|
size of the real RSB in hardware (which has 3 entries). Larger RSBs than 8
|
||||||
|
showed degraded performance.
|
||||||
|
|
||||||
|
struct JitState {
|
||||||
|
// ...
|
||||||
|
|
||||||
|
static constexpr size_t RSBSize = 8; // MUST be a power of 2.
|
||||||
|
u32 rsb_ptr = 0;
|
||||||
|
std::array<u64, RSBSize> rsb_location_descriptors;
|
||||||
|
std::array<u64, RSBSize> rsb_codeptrs;
|
||||||
|
void ResetRSB();
|
||||||
|
|
||||||
|
// ...
|
||||||
|
};
|
||||||
|
|
||||||
|
### RSB Push
|
||||||
|
|
||||||
|
We insert our prediction at the insertion point iff the RSB doesn't already
|
||||||
|
contain a prediction with the same `UniqueHash`.
|
||||||
|
|
||||||
|
void EmitX64::EmitPushRSB(IR::Block&, IR::Inst* inst) {
|
||||||
|
using namespace Xbyak::util;
|
||||||
|
|
||||||
|
ASSERT(inst->GetArg(0).IsImmediate());
|
||||||
|
u64 imm64 = inst->GetArg(0).GetU64();
|
||||||
|
|
||||||
|
Xbyak::Reg64 code_ptr_reg = reg_alloc.ScratchGpr({HostLoc::RCX});
|
||||||
|
Xbyak::Reg64 loc_desc_reg = reg_alloc.ScratchGpr();
|
||||||
|
Xbyak::Reg32 index_reg = reg_alloc.ScratchGpr().cvt32();
|
||||||
|
u64 code_ptr = unique_hash_to_code_ptr.find(imm64) != unique_hash_to_code_ptr.end()
|
||||||
|
? u64(unique_hash_to_code_ptr[imm64])
|
||||||
|
: u64(code->GetReturnFromRunCodeAddress());
|
||||||
|
|
||||||
|
code->mov(index_reg, dword[r15 + offsetof(JitState, rsb_ptr)]);
|
||||||
|
code->add(index_reg, 1);
|
||||||
|
code->and_(index_reg, u32(JitState::RSBSize - 1));
|
||||||
|
|
||||||
|
code->mov(loc_desc_reg, u64(imm64));
|
||||||
|
CodePtr patch_location = code->getCurr<CodePtr>();
|
||||||
|
patch_unique_hash_locations[imm64].emplace_back(patch_location);
|
||||||
|
code->mov(code_ptr_reg, u64(code_ptr)); // This line has to match up with EmitX64::Patch.
|
||||||
|
code->EnsurePatchLocationSize(patch_location, 10);
|
||||||
|
|
||||||
|
Xbyak::Label label;
|
||||||
|
for (size_t i = 0; i < JitState::RSBSize; ++i) {
|
||||||
|
code->cmp(loc_desc_reg, qword[r15 + offsetof(JitState, rsb_location_descriptors) + i * sizeof(u64)]);
|
||||||
|
code->je(label, code->T_SHORT);
|
||||||
|
}
|
||||||
|
|
||||||
|
code->mov(dword[r15 + offsetof(JitState, rsb_ptr)], index_reg);
|
||||||
|
code->mov(qword[r15 + index_reg.cvt64() * 8 + offsetof(JitState, rsb_location_descriptors)], loc_desc_reg);
|
||||||
|
code->mov(qword[r15 + index_reg.cvt64() * 8 + offsetof(JitState, rsb_codeptrs)], code_ptr_reg);
|
||||||
|
code->L(label);
|
||||||
|
}
|
||||||
|
|
||||||
|
In pseudocode:
|
||||||
|
|
||||||
|
for (i := 0 .. RSBSize-1)
|
||||||
|
if (rsb_location_descriptors[i] == imm64)
|
||||||
|
goto label;
|
||||||
|
rsb_ptr++;
|
||||||
|
rsb_ptr %= RSBSize;
|
||||||
|
rsb_location_desciptors[rsb_ptr] = imm64; //< The UniqueHash
|
||||||
|
rsb_codeptr[rsb_ptr] = /* codeptr corresponding to the UniqueHash */;
|
||||||
|
label:
|
||||||
|
|
||||||
|
## RSB Pop
|
||||||
|
|
||||||
|
To check if a predicition is in the RSB, we linearly scan the RSB.
|
||||||
|
|
||||||
|
void EmitX64::EmitTerminalPopRSBHint(IR::Term::PopRSBHint, IR::LocationDescriptor initial_location) {
|
||||||
|
using namespace Xbyak::util;
|
||||||
|
|
||||||
|
// This calculation has to match up with IREmitter::PushRSB
|
||||||
|
code->mov(ebx, MJitStateCpsr());
|
||||||
|
code->mov(ecx, MJitStateReg(Arm::Reg::PC));
|
||||||
|
code->and_(ebx, u32((1 << 5) | (1 << 9)));
|
||||||
|
code->shr(ebx, 2);
|
||||||
|
code->or_(ebx, dword[r15 + offsetof(JitState, FPSCR_mode)]);
|
||||||
|
code->shl(rbx, 32);
|
||||||
|
code->or_(rbx, rcx);
|
||||||
|
|
||||||
|
code->mov(rax, u64(code->GetReturnFromRunCodeAddress()));
|
||||||
|
for (size_t i = 0; i < JitState::RSBSize; ++i) {
|
||||||
|
code->cmp(rbx, qword[r15 + offsetof(JitState, rsb_location_descriptors) + i * sizeof(u64)]);
|
||||||
|
code->cmove(rax, qword[r15 + offsetof(JitState, rsb_codeptrs) + i * sizeof(u64)]);
|
||||||
|
}
|
||||||
|
|
||||||
|
code->jmp(rax);
|
||||||
|
}
|
||||||
|
|
||||||
|
In pseudocode:
|
||||||
|
|
||||||
|
rbx := ComputeUniqueHash()
|
||||||
|
rax := ReturnToDispatch
|
||||||
|
for (i := 0 .. RSBSize-1)
|
||||||
|
if (rbx == rsb_location_descriptors[i])
|
||||||
|
rax = rsb_codeptrs[i]
|
||||||
|
goto rax
|
||||||
Loading…
Reference in a new issue