fabi/Xenia-Canary
1
0
Fork 0
Code Issues Pull Requests Actions Packages Projects Releases Wiki Activity

Minor decoder optimizations, kernel fixes, cpu backend fixes

Browse Source
This commit is contained in:
chss95cs@gmail.com
2022-11-05 10:50:33 -07:00
parent ba66373d8c
commit c1d922eebf
62 changed files with 1254 additions and 802 deletions
Download Patch File Download Diff File Expand all files Collapse all files

3
src/xenia/cpu/backend/x64/x64_code_cache.h
View File

@@ -93,7 +93,8 @@ class X64CodeCache : public CodeCache {
// This is picked to be high enough to cover whatever we can reasonably
// expect. If we hit issues with this it probably means some corner case
// in analysis triggering.
static const size_t kMaximumFunctionCount = 100000;
//chrispy: raised this, some games that were compiled with low optimization levels can exceed this
static const size_t kMaximumFunctionCount = 1000000;
struct UnwindReservation {
size_t data_size = 0;

9
src/xenia/cpu/backend/x64/x64_code_cache_win.cc
View File

@@ -209,7 +209,16 @@ bool Win32X64CodeCache::Initialize() {
Win32X64CodeCache::UnwindReservation
Win32X64CodeCache::RequestUnwindReservation(uint8_t* entry_address) {
#if defined(NDEBUG)
if (unwind_table_count_ >= kMaximumFunctionCount) {
// we should not just be ignoring this in release if it happens
xe::FatalError(
"Unwind table count (unwind_table_count_) exceeded maximum! Please report this to "
"Xenia/Canary developers");
}
#else
assert_false(unwind_table_count_ >= kMaximumFunctionCount);
#endif
UnwindReservation unwind_reservation;
unwind_reservation.data_size = xe::round_up(kUnwindInfoSize, 16);
unwind_reservation.table_slot = unwind_table_count_++;

96
src/xenia/cpu/backend/x64/x64_emitter.cc
View File

@@ -46,10 +46,6 @@ DEFINE_bool(ignore_undefined_externs, true,
DEFINE_bool(emit_source_annotations, false,
"Add extra movs and nops to make disassembly easier to read.",
"CPU");
DEFINE_bool(resolve_rel32_guest_calls, true,
"Experimental optimization, directly call already resolved "
"functions via x86 rel32 call/jmp",
"CPU");
DEFINE_bool(enable_incorrect_roundingmode_behavior, false,
"Disables the FPU/VMX MXCSR sharing workaround, potentially "
@@ -78,7 +74,6 @@ using namespace xe::literals;
static const size_t kMaxCodeSize = 1_MiB;
static const size_t kStashOffset = 32;
// static const size_t kStashOffsetHigh = 32 + 32;
const uint32_t X64Emitter::gpr_reg_map_[X64Emitter::GPR_COUNT] = {
@@ -141,55 +136,6 @@ bool X64Emitter::Emit(GuestFunction* function, HIRBuilder* builder,
return true;
}
#pragma pack(push, 1)
struct RGCEmitted {
uint8_t ff_;
uint32_t rgcid_;
};
#pragma pack(pop)
#if 0
void X64Emitter::InjectCallAddresses(void* new_execute_address) {
for (auto&& callsite : call_sites_) {
RGCEmitted* hunter = (RGCEmitted*)new_execute_address;
while (hunter->ff_ != 0xFF || hunter->rgcid_ != callsite.offset_) {
hunter =
reinterpret_cast<RGCEmitted*>(reinterpret_cast<char*>(hunter) + 1);
}
hunter->ff_ = callsite.is_jump_ ? 0xE9 : 0xE8;
hunter->rgcid_ =
static_cast<uint32_t>(static_cast<intptr_t>(callsite.destination_) -
reinterpret_cast<intptr_t>(hunter + 1));
}
}
#else
void X64Emitter::InjectCallAddresses(void* new_execute_address) {
#if 0
RGCEmitted* hunter = (RGCEmitted*)new_execute_address;
std::map<uint32_t, ResolvableGuestCall*> id_to_rgc{};
for (auto&& callsite : call_sites_) {
id_to_rgc[callsite.offset_] = &callsite;
}
#else
RGCEmitted* hunter = (RGCEmitted*)new_execute_address;
for (auto&& callsite : call_sites_) {
while (hunter->ff_ != 0xFF || hunter->rgcid_ != callsite.offset_) {
hunter =
reinterpret_cast<RGCEmitted*>(reinterpret_cast<char*>(hunter) + 1);
}
hunter->ff_ = callsite.is_jump_ ? 0xE9 : 0xE8;
hunter->rgcid_ =
static_cast<uint32_t>(static_cast<intptr_t>(callsite.destination_) -
reinterpret_cast<intptr_t>(hunter + 1));
}
#endif
}
#endif
void* X64Emitter::Emplace(const EmitFunctionInfo& func_info,
GuestFunction* function) {
// To avoid changing xbyak, we do a switcharoo here.
@@ -207,10 +153,6 @@ void* X64Emitter::Emplace(const EmitFunctionInfo& func_info,
if (function) {
code_cache_->PlaceGuestCode(function->address(), top_, func_info, function,
new_execute_address, new_write_address);
if (cvars::resolve_rel32_guest_calls) {
InjectCallAddresses(new_execute_address);
}
} else {
code_cache_->PlaceHostCode(0, top_, func_info, new_execute_address,
new_write_address);
@@ -219,7 +161,6 @@ void* X64Emitter::Emplace(const EmitFunctionInfo& func_info,
ready();
top_ = old_address;
reset();
call_sites_.clear();
tail_code_.clear();
for (auto&& cached_label : label_cache_) {
delete cached_label;
@@ -336,7 +277,7 @@ bool X64Emitter::Emit(HIRBuilder* builder, EmitFunctionInfo& func_info) {
// Mark block labels.
auto label = block->label_head;
while (label) {
L(label->name);
L(std::to_string(label->id));
label = label->next;
}
@@ -418,7 +359,6 @@ void X64Emitter::EmitProfilerEpilogue() {
// actually... lets just try without atomics lol
// lock();
add(qword[r10], rdx);
}
#endif
}
@@ -534,44 +474,23 @@ void X64Emitter::Call(const hir::Instr* instr, GuestFunction* function) {
auto fn = static_cast<X64Function*>(function);
// Resolve address to the function to call and store in rax.
if (cvars::resolve_rel32_guest_calls && fn->machine_code()) {
ResolvableGuestCall rgc;
rgc.destination_ = uint32_t(uint64_t(fn->machine_code()));
rgc.offset_ = current_rgc_id_;
current_rgc_id_++;
if (fn->machine_code()) {
if (!(instr->flags & hir::CALL_TAIL)) {
mov(rcx, qword[rsp + StackLayout::GUEST_CALL_RET_ADDR]);
db(0xFF);
rgc.is_jump_ = false;
dd(rgc.offset_);
call((void*)fn->machine_code());
} else {
// tail call
EmitTraceUserCallReturn();
rgc.is_jump_ = true;
EmitProfilerEpilogue();
// Pass the callers return address over.
mov(rcx, qword[rsp + StackLayout::GUEST_RET_ADDR]);
add(rsp, static_cast<uint32_t>(stack_size()));
db(0xFF);
dd(rgc.offset_);
jmp((void*)fn->machine_code(), T_NEAR);
}
call_sites_.push_back(rgc);
return;
}
if (fn->machine_code()) {
// TODO(benvanik): is it worth it to do this? It removes the need for
// a ResolveFunction call, but makes the table less useful.
assert_zero(uint64_t(fn->machine_code()) & 0xFFFFFFFF00000000);
// todo: this should be changed so that we can actually do a call to
// fn->machine_code. the code will be emitted near us, so 32 bit rel jmp
// should be possible
mov(eax, uint32_t(uint64_t(fn->machine_code())));
} else if (code_cache_->has_indirection_table()) {
// Load the pointer to the indirection table maintained in X64CodeCache.
// The target dword will either contain the address of the generated code
@@ -1017,7 +936,10 @@ static const vec128_t xmm_consts[] = {
/*XMMSTVLShuffle*/
v128_setr_bytes(0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15),
/* XMMSTVRSwapMask*/
vec128b((uint8_t)0x83)};
vec128b((uint8_t)0x83), /*XMMVSRShlByteshuf*/
v128_setr_bytes(13, 14, 15, 8, 9, 10, 11, 4, 5, 6, 7, 0, 1, 2, 3, 0x80),
// XMMVSRMask
vec128b(1)};
void* X64Emitter::FindByteConstantOffset(unsigned bytevalue) {
for (auto& vec : xmm_consts) {

18
src/xenia/cpu/backend/x64/x64_emitter.h
View File

@@ -66,7 +66,7 @@ enum class SimdDomain : uint32_t {
};
enum class MXCSRMode : uint32_t { Unknown, Fpu, Vmx };
XE_MAYBE_UNUSED
static SimdDomain PickDomain2(SimdDomain dom1, SimdDomain dom2) {
if (dom1 == dom2) {
return dom1;
@@ -172,7 +172,9 @@ enum XmmConst {
XMMLVLShuffle,
XMMLVRCmp16,
XMMSTVLShuffle,
XMMSTVRSwapMask // swapwordmask with bit 7 set
XMMSTVRSwapMask, // swapwordmask with bit 7 set
XMMVSRShlByteshuf,
XMMVSRMask
};
using amdfx::xopcompare_e;
@@ -190,13 +192,6 @@ class XbyakAllocator : public Xbyak::Allocator {
virtual bool useProtect() const { return false; }
};
class ResolvableGuestCall {
public:
bool is_jump_;
uintptr_t destination_;
// rgcid
unsigned offset_;
};
class X64Emitter;
using TailEmitCallback = std::function<void(X64Emitter& e, Xbyak::Label& lbl)>;
struct TailEmitter {
@@ -220,7 +215,6 @@ class X64Emitter : public Xbyak::CodeGenerator {
uint32_t debug_info_flags, FunctionDebugInfo* debug_info,
void** out_code_address, size_t* out_code_size,
std::vector<SourceMapEntry>* out_source_map);
void InjectCallAddresses(void* new_execute_addr);
public:
// Reserved: rsp, rsi, rdi
@@ -230,7 +224,7 @@ class X64Emitter : public Xbyak::CodeGenerator {
// xmm4-xmm15 (save to get xmm3)
static const int GPR_COUNT = 7;
static const int XMM_COUNT = 12;
static constexpr size_t kStashOffset = 32;
static void SetupReg(const hir::Value* v, Xbyak::Reg8& r) {
auto idx = gpr_reg_map_[v->reg.index];
r = Xbyak::Reg8(idx);
@@ -410,8 +404,6 @@ class X64Emitter : public Xbyak::CodeGenerator {
static const uint32_t gpr_reg_map_[GPR_COUNT];
static const uint32_t xmm_reg_map_[XMM_COUNT];
uint32_t current_rgc_id_ = 0xEEDDF00F;
std::vector<ResolvableGuestCall> call_sites_;
/*
set to true if the low 32 bits of membase == 0.
only really advantageous if you are storing 32 bit 0 to a displaced address,

9
src/xenia/cpu/backend/x64/x64_op.h
View File

@@ -398,21 +398,22 @@ struct I<OPCODE, DEST, SRC1, SRC2, SRC3> : DestField<DEST> {
};
template <typename T>
XE_MAYBE_UNUSED
static const T GetTempReg(X64Emitter& e);
template <>
const Reg8 GetTempReg<Reg8>(X64Emitter& e) {
XE_MAYBE_UNUSED const Reg8 GetTempReg<Reg8>(X64Emitter& e) {
return e.al;
}
template <>
const Reg16 GetTempReg<Reg16>(X64Emitter& e) {
XE_MAYBE_UNUSED const Reg16 GetTempReg<Reg16>(X64Emitter& e) {
return e.ax;
}
template <>
const Reg32 GetTempReg<Reg32>(X64Emitter& e) {
XE_MAYBE_UNUSED const Reg32 GetTempReg<Reg32>(X64Emitter& e) {
return e.eax;
}
template <>
const Reg64 GetTempReg<Reg64>(X64Emitter& e) {
XE_MAYBE_UNUSED const Reg64 GetTempReg<Reg64>(X64Emitter& e) {
return e.rax;
}

44
src/xenia/cpu/backend/x64/x64_seq_control.cc
View File

@@ -25,46 +25,46 @@ static void EmitFusedBranch(X64Emitter& e, const T& i) {
bool valid = i.instr->prev && i.instr->prev->dest == i.src1.value;
auto opcode = valid ? i.instr->prev->opcode->num : -1;
if (valid) {
auto name = i.src2.value->name;
std::string name = i.src2.value->GetIdString();
switch (opcode) {
case OPCODE_COMPARE_EQ:
e.je(name, e.T_NEAR);
e.je(std::move(name), e.T_NEAR);
break;
case OPCODE_COMPARE_NE:
e.jne(name, e.T_NEAR);
e.jne(std::move(name), e.T_NEAR);
break;
case OPCODE_COMPARE_SLT:
e.jl(name, e.T_NEAR);
e.jl(std::move(name), e.T_NEAR);
break;
case OPCODE_COMPARE_SLE:
e.jle(name, e.T_NEAR);
e.jle(std::move(name), e.T_NEAR);
break;
case OPCODE_COMPARE_SGT:
e.jg(name, e.T_NEAR);
e.jg(std::move(name), e.T_NEAR);
break;
case OPCODE_COMPARE_SGE:
e.jge(name, e.T_NEAR);
e.jge(std::move(name), e.T_NEAR);
break;
case OPCODE_COMPARE_ULT:
e.jb(name, e.T_NEAR);
e.jb(std::move(name), e.T_NEAR);
break;
case OPCODE_COMPARE_ULE:
e.jbe(name, e.T_NEAR);
e.jbe(std::move(name), e.T_NEAR);
break;
case OPCODE_COMPARE_UGT:
e.ja(name, e.T_NEAR);
e.ja(std::move(name), e.T_NEAR);
break;
case OPCODE_COMPARE_UGE:
e.jae(name, e.T_NEAR);
e.jae(std::move(name), e.T_NEAR);
break;
default:
e.test(i.src1, i.src1);
e.jnz(name, e.T_NEAR);
e.jnz(std::move(name), e.T_NEAR);
break;
}
} else {
e.test(i.src1, i.src1);
e.jnz(i.src2.value->name, e.T_NEAR);
e.jnz(i.src2.value->GetIdString(), e.T_NEAR);
}
}
// ============================================================================
@@ -490,7 +490,7 @@ EMITTER_OPCODE_TABLE(OPCODE_SET_RETURN_ADDRESS, SET_RETURN_ADDRESS);
// ============================================================================
struct BRANCH : Sequence<BRANCH, I<OPCODE_BRANCH, VoidOp, LabelOp>> {
static void Emit(X64Emitter& e, const EmitArgType& i) {
e.jmp(i.src1.value->name, e.T_NEAR);
e.jmp(i.src1.value->GetIdString(), e.T_NEAR);
}
};
EMITTER_OPCODE_TABLE(OPCODE_BRANCH, BRANCH);
@@ -534,7 +534,7 @@ struct BRANCH_TRUE_F32
Xmm input = GetInputRegOrConstant(e, i.src1, e.xmm0);
e.vmovd(e.eax, input);
e.test(e.eax, e.eax);
e.jnz(i.src2.value->name, e.T_NEAR);
e.jnz(i.src2.value->GetIdString(), e.T_NEAR);
}
};
struct BRANCH_TRUE_F64
@@ -543,7 +543,7 @@ struct BRANCH_TRUE_F64
Xmm input = GetInputRegOrConstant(e, i.src1, e.xmm0);
e.vmovq(e.rax, input);
e.test(e.rax, e.rax);
e.jnz(i.src2.value->name, e.T_NEAR);
e.jnz(i.src2.value->GetIdString(), e.T_NEAR);
}
};
EMITTER_OPCODE_TABLE(OPCODE_BRANCH_TRUE, BRANCH_TRUE_I8, BRANCH_TRUE_I16,
@@ -557,7 +557,7 @@ struct BRANCH_FALSE_I8
: Sequence<BRANCH_FALSE_I8, I<OPCODE_BRANCH_FALSE, VoidOp, I8Op, LabelOp>> {
static void Emit(X64Emitter& e, const EmitArgType& i) {
e.test(i.src1, i.src1);
e.jz(i.src2.value->name, e.T_NEAR);
e.jz(i.src2.value->GetIdString(), e.T_NEAR);
}
};
struct BRANCH_FALSE_I16
@@ -565,7 +565,7 @@ struct BRANCH_FALSE_I16
I<OPCODE_BRANCH_FALSE, VoidOp, I16Op, LabelOp>> {
static void Emit(X64Emitter& e, const EmitArgType& i) {
e.test(i.src1, i.src1);
e.jz(i.src2.value->name, e.T_NEAR);
e.jz(i.src2.value->GetIdString(), e.T_NEAR);
}
};
struct BRANCH_FALSE_I32
@@ -573,7 +573,7 @@ struct BRANCH_FALSE_I32
I<OPCODE_BRANCH_FALSE, VoidOp, I32Op, LabelOp>> {
static void Emit(X64Emitter& e, const EmitArgType& i) {
e.test(i.src1, i.src1);
e.jz(i.src2.value->name, e.T_NEAR);
e.jz(i.src2.value->GetIdString(), e.T_NEAR);
}
};
struct BRANCH_FALSE_I64
@@ -581,7 +581,7 @@ struct BRANCH_FALSE_I64
I<OPCODE_BRANCH_FALSE, VoidOp, I64Op, LabelOp>> {
static void Emit(X64Emitter& e, const EmitArgType& i) {
e.test(i.src1, i.src1);
e.jz(i.src2.value->name, e.T_NEAR);
e.jz(i.src2.value->GetIdString(), e.T_NEAR);
}
};
struct BRANCH_FALSE_F32
@@ -591,7 +591,7 @@ struct BRANCH_FALSE_F32
Xmm input = GetInputRegOrConstant(e, i.src1, e.xmm0);
e.vmovd(e.eax, input);
e.test(e.eax, e.eax);
e.jz(i.src2.value->name, e.T_NEAR);
e.jz(i.src2.value->GetIdString(), e.T_NEAR);
}
};
struct BRANCH_FALSE_F64
@@ -601,7 +601,7 @@ struct BRANCH_FALSE_F64
Xmm input = GetInputRegOrConstant(e, i.src1, e.xmm0);
e.vmovq(e.rax, input);
e.test(e.rax, e.rax);
e.jz(i.src2.value->name, e.T_NEAR);
e.jz(i.src2.value->GetIdString(), e.T_NEAR);
}
};
EMITTER_OPCODE_TABLE(OPCODE_BRANCH_FALSE, BRANCH_FALSE_I8, BRANCH_FALSE_I16,

778
src/xenia/cpu/backend/x64/x64_seq_vector.cc
View File

@@ -805,22 +805,7 @@ EMITTER_OPCODE_TABLE(OPCODE_VECTOR_SUB, VECTOR_SUB);
// ============================================================================
// OPCODE_VECTOR_SHL
// ============================================================================
template <typename T, std::enable_if_t<std::is_integral<T>::value, int> = 0>
static __m128i EmulateVectorShl(void*, __m128i src1, __m128i src2) {
alignas(16) T value[16 / sizeof(T)];
alignas(16) T shamt[16 / sizeof(T)];
// Load SSE registers into a C array.
_mm_store_si128(reinterpret_cast<__m128i*>(value), src1);
_mm_store_si128(reinterpret_cast<__m128i*>(shamt), src2);
for (size_t i = 0; i < (16 / sizeof(T)); ++i) {
value[i] = value[i] << (shamt[i] & ((sizeof(T) * 8) - 1));
}
// Store result and return it.
return _mm_load_si128(reinterpret_cast<__m128i*>(value));
}
static XmmConst GetShiftmaskForType(unsigned typ) {
if (typ == INT8_TYPE) {
return XMMXOPByteShiftMask;
@@ -914,28 +899,14 @@ struct VECTOR_SHL_V128
}
}
if (all_same) {
// mul by two
/*if (seenvalue == 1) {
e.vpaddb(i.dest, i.src1, i.src1);
} else if (seenvalue == 2) {
e.vpaddb(i.dest, i.src1, i.src1);
e.vpaddb(i.dest, i.dest, i.dest);
} else if (seenvalue == 3) {
// mul by 8
e.vpaddb(i.dest, i.src1, i.src1);
e.vpaddb(i.dest, i.dest, i.dest);
e.vpaddb(i.dest, i.dest, i.dest);
} else*/
{
e.vpmovzxbw(e.ymm0, i.src1);
e.vpsllw(e.ymm0, e.ymm0, seenvalue);
e.vextracti128(e.xmm1, e.ymm0, 1);
e.vpmovzxbw(e.ymm0, i.src1);
e.vpsllw(e.ymm0, e.ymm0, seenvalue);
e.vextracti128(e.xmm1, e.ymm0, 1);
e.vpshufb(e.xmm0, e.xmm0, e.GetXmmConstPtr(XMMShortsToBytes));
e.vpshufb(e.xmm1, e.xmm1, e.GetXmmConstPtr(XMMShortsToBytes));
e.vpunpcklqdq(i.dest, e.xmm0, e.xmm1);
return;
}
e.vpshufb(e.xmm0, e.xmm0, e.GetXmmConstPtr(XMMShortsToBytes));
e.vpshufb(e.xmm1, e.xmm1, e.GetXmmConstPtr(XMMShortsToBytes));
e.vpunpcklqdq(i.dest, e.xmm0, e.xmm1);
return;
} else {
e.LoadConstantXmm(e.xmm2, constmask);
@@ -966,14 +937,41 @@ struct VECTOR_SHL_V128
}
}
}
if (i.src2.is_constant) {
e.lea(e.GetNativeParam(1), e.StashConstantXmm(1, i.src2.constant()));
unsigned stack_offset_src1 = StackLayout::GUEST_SCRATCH;
unsigned stack_offset_src2 = StackLayout::GUEST_SCRATCH + 16;
if (i.src1.is_constant) {
e.StashConstantXmm(0, i.src1.constant());
stack_offset_src1 = X64Emitter::kStashOffset;
} else {
e.lea(e.GetNativeParam(1), e.StashXmm(1, i.src2));
e.vmovdqa(e.ptr[e.rsp + stack_offset_src1], i.src1);
}
e.lea(e.GetNativeParam(0), e.StashXmm(0, i.src1));
e.CallNativeSafe(reinterpret_cast<void*>(EmulateVectorShl<uint8_t>));
e.vmovaps(i.dest, e.xmm0);
if (i.src2.is_constant) {
e.StashConstantXmm(1, i.src2.constant());
stack_offset_src2 = X64Emitter::kStashOffset + 16;
} else {
e.vmovdqa(e.ptr[e.rsp + stack_offset_src2], i.src2);
}
Xbyak::Label looper;
e.xor_(e.edx, e.edx);
e.L(looper);
e.movzx(e.ecx, e.byte[e.rsp + stack_offset_src2 + e.rdx]);
e.shl(e.byte[e.rsp + stack_offset_src1 + e.rdx], e.cl);
if (e.IsFeatureEnabled(kX64FlagsIndependentVars)) {
e.inc(e.edx);
} else {
e.add(e.edx, 1);
}
e.cmp(e.edx, 16);
e.jnz(looper);
e.vmovdqa(i.dest, e.byte[e.rsp + stack_offset_src1]);
}
static void EmitInt16(X64Emitter& e, const EmitArgType& i) {
Xmm src1;
@@ -1022,14 +1020,32 @@ struct VECTOR_SHL_V128
// TODO(benvanik): native version (with shift magic).
e.L(emu);
unsigned stack_offset_src1 = StackLayout::GUEST_SCRATCH;
unsigned stack_offset_src2 = StackLayout::GUEST_SCRATCH + 16;
e.vmovdqa(e.ptr[e.rsp + stack_offset_src1], src1);
if (i.src2.is_constant) {
e.lea(e.GetNativeParam(1), e.StashConstantXmm(1, i.src2.constant()));
e.StashConstantXmm(1, i.src2.constant());
stack_offset_src2 = X64Emitter::kStashOffset + 16;
} else {
e.lea(e.GetNativeParam(1), e.StashXmm(1, i.src2));
e.vmovdqa(e.ptr[e.rsp + stack_offset_src2], i.src2);
}
e.lea(e.GetNativeParam(0), e.StashXmm(0, src1));
e.CallNativeSafe(reinterpret_cast<void*>(EmulateVectorShl<uint16_t>));
e.vmovaps(i.dest, e.xmm0);
Xbyak::Label looper;
e.xor_(e.edx, e.edx);
e.L(looper);
e.movzx(e.ecx, e.word[e.rsp + stack_offset_src2 + e.rdx]);
e.shl(e.word[e.rsp + stack_offset_src1 + e.rdx], e.cl);
e.add(e.edx, 2);
e.cmp(e.edx, 16);
e.jnz(looper);
e.vmovdqa(i.dest, e.byte[e.rsp + stack_offset_src1]);
e.L(end);
}
@@ -1098,14 +1114,32 @@ struct VECTOR_SHL_V128
// TODO(benvanik): native version (with shift magic).
e.L(emu);
unsigned stack_offset_src1 = StackLayout::GUEST_SCRATCH;
unsigned stack_offset_src2 = StackLayout::GUEST_SCRATCH + 16;
e.vmovdqa(e.ptr[e.rsp + stack_offset_src1], src1);
if (i.src2.is_constant) {
e.lea(e.GetNativeParam(1), e.StashConstantXmm(1, i.src2.constant()));
e.StashConstantXmm(1, i.src2.constant());
stack_offset_src2 = X64Emitter::kStashOffset + 16;
} else {
e.lea(e.GetNativeParam(1), e.StashXmm(1, i.src2));
e.vmovdqa(e.ptr[e.rsp + stack_offset_src2], i.src2);
}
e.lea(e.GetNativeParam(0), e.StashXmm(0, src1));
e.CallNativeSafe(reinterpret_cast<void*>(EmulateVectorShl<uint32_t>));
e.vmovaps(i.dest, e.xmm0);
Xbyak::Label looper;
e.xor_(e.edx, e.edx);
e.L(looper);
e.mov(e.ecx, e.dword[e.rsp + stack_offset_src2 + e.rdx]);
e.shl(e.dword[e.rsp + stack_offset_src1 + e.rdx], e.cl);
e.add(e.edx, 4);
e.cmp(e.edx, 16);
e.jnz(looper);
e.vmovdqa(i.dest, e.byte[e.rsp + stack_offset_src1]);
e.L(end);
}
@@ -1116,22 +1150,6 @@ EMITTER_OPCODE_TABLE(OPCODE_VECTOR_SHL, VECTOR_SHL_V128);
// ============================================================================
// OPCODE_VECTOR_SHR
// ============================================================================
template <typename T, std::enable_if_t<std::is_integral<T>::value, int> = 0>
static __m128i EmulateVectorShr(void*, __m128i src1, __m128i src2) {
alignas(16) T value[16 / sizeof(T)];
alignas(16) T shamt[16 / sizeof(T)];
// Load SSE registers into a C array.
_mm_store_si128(reinterpret_cast<__m128i*>(value), src1);
_mm_store_si128(reinterpret_cast<__m128i*>(shamt), src2);
for (size_t i = 0; i < (16 / sizeof(T)); ++i) {
value[i] = value[i] >> (shamt[i] & ((sizeof(T) * 8) - 1));
}
// Store result and return it.
return _mm_load_si128(reinterpret_cast<__m128i*>(value));
}
struct VECTOR_SHR_V128
: Sequence<VECTOR_SHR_V128, I<OPCODE_VECTOR_SHR, V128Op, V128Op, V128Op>> {
@@ -1179,34 +1197,63 @@ struct VECTOR_SHR_V128
}
static void EmitInt8(X64Emitter& e, const EmitArgType& i) {
// TODO(benvanik): native version (with shift magic).
if (i.src2.is_constant) {
if (e.IsFeatureEnabled(kX64EmitGFNI)) {
const auto& shamt = i.src2.constant();
bool all_same = true;
for (size_t n = 0; n < 16 - n; ++n) {
if (shamt.u8[n] != shamt.u8[n + 1]) {
all_same = false;
break;
}
}
if (all_same) {
// Every count is the same, so we can use gf2p8affineqb.
const uint8_t shift_amount = shamt.u8[0] & 0b111;
const uint64_t shift_matrix = UINT64_C(0x0102040810204080)
<< (shift_amount * 8);
e.vgf2p8affineqb(i.dest, i.src1,
e.StashConstantXmm(0, vec128q(shift_matrix)), 0);
return;
if (i.src2.is_constant && e.IsFeatureEnabled(kX64EmitGFNI)) {
const auto& shamt = i.src2.constant();
bool all_same = true;
for (size_t n = 0; n < 16 - n; ++n) {
if (shamt.u8[n] != shamt.u8[n + 1]) {
all_same = false;
break;
}
}
e.lea(e.GetNativeParam(1), e.StashConstantXmm(1, i.src2.constant()));
} else {
e.lea(e.GetNativeParam(1), e.StashXmm(1, i.src2));
if (all_same) {
// Every count is the same, so we can use gf2p8affineqb.
const uint8_t shift_amount = shamt.u8[0] & 0b111;
const uint64_t shift_matrix = UINT64_C(0x0102040810204080)
<< (shift_amount * 8);
e.vgf2p8affineqb(i.dest, i.src1,
e.StashConstantXmm(0, vec128q(shift_matrix)), 0);
return;
}
}
e.lea(e.GetNativeParam(0), e.StashXmm(0, i.src1));
e.CallNativeSafe(reinterpret_cast<void*>(EmulateVectorShr<uint8_t>));
e.vmovaps(i.dest, e.xmm0);
unsigned stack_offset_src1 = StackLayout::GUEST_SCRATCH;
unsigned stack_offset_src2 = StackLayout::GUEST_SCRATCH + 16;
if (i.src1.is_constant) {
e.StashConstantXmm(0, i.src1.constant());
stack_offset_src1 = X64Emitter::kStashOffset;
} else {
e.vmovdqa(e.ptr[e.rsp + stack_offset_src1], i.src1);
}
if (i.src2.is_constant) {
e.StashConstantXmm(1, i.src2.constant());
stack_offset_src2 = X64Emitter::kStashOffset + 16;
} else {
e.vmovdqa(e.ptr[e.rsp + stack_offset_src2], i.src2);
}
Xbyak::Label looper;
e.xor_(e.edx, e.edx);
e.L(looper);
// movzx is to eliminate any possible dep on previous value of rcx at start
// of loop
e.movzx(e.ecx, e.byte[e.rsp + stack_offset_src2 + e.rdx]);
// maybe using a memory operand as the left side isn't the best idea lol,
// still better than callnativesafe though agners docs have no timing info
// on shx [m], cl so shrug
e.shr(e.byte[e.rsp + stack_offset_src1 + e.rdx], e.cl);
if (e.IsFeatureEnabled(kX64FlagsIndependentVars)) {
e.inc(e.edx);
} else {
e.add(e.edx, 1);
}
e.cmp(e.edx, 16);
e.jnz(looper);
e.vmovdqa(i.dest, e.byte[e.rsp + stack_offset_src1]);
}
static void EmitInt16(X64Emitter& e, const EmitArgType& i) {
@@ -1248,14 +1295,38 @@ struct VECTOR_SHR_V128
// TODO(benvanik): native version (with shift magic).
e.L(emu);
if (i.src2.is_constant) {
e.lea(e.GetNativeParam(1), e.StashConstantXmm(1, i.src2.constant()));
unsigned stack_offset_src1 = StackLayout::GUEST_SCRATCH;
unsigned stack_offset_src2 = StackLayout::GUEST_SCRATCH + 16;
if (i.src1.is_constant) {
e.StashConstantXmm(0, i.src1.constant());
stack_offset_src1 = X64Emitter::kStashOffset;
} else {
e.lea(e.GetNativeParam(1), e.StashXmm(1, i.src2));
e.vmovdqa(e.ptr[e.rsp + stack_offset_src1], i.src1);
}
e.lea(e.GetNativeParam(0), e.StashXmm(0, i.src1));
e.CallNativeSafe(reinterpret_cast<void*>(EmulateVectorShr<uint16_t>));
e.vmovaps(i.dest, e.xmm0);
if (i.src2.is_constant) {
e.StashConstantXmm(1, i.src2.constant());
stack_offset_src2 = X64Emitter::kStashOffset + 16;
} else {
e.vmovdqa(e.ptr[e.rsp + stack_offset_src2], i.src2);
}
Xbyak::Label looper;
e.xor_(e.edx, e.edx);
e.L(looper);
e.movzx(e.ecx, e.word[e.rsp + stack_offset_src2 + e.rdx]);
e.shr(e.word[e.rsp + stack_offset_src1 + e.rdx], e.cl);
e.add(e.edx, 2);
e.cmp(e.edx, 16);
e.jnz(looper);
e.vmovdqa(i.dest, e.byte[e.rsp + stack_offset_src1]);
e.L(end);
}
@@ -1324,14 +1395,37 @@ struct VECTOR_SHR_V128
// TODO(benvanik): native version.
e.L(emu);
if (i.src2.is_constant) {
e.lea(e.GetNativeParam(1), e.StashConstantXmm(1, i.src2.constant()));
unsigned stack_offset_src1 = StackLayout::GUEST_SCRATCH;
unsigned stack_offset_src2 = StackLayout::GUEST_SCRATCH + 16;
if (i.src1.is_constant) {
e.StashConstantXmm(0, i.src1.constant());
stack_offset_src1 = X64Emitter::kStashOffset;
} else {
e.lea(e.GetNativeParam(1), e.StashXmm(1, i.src2));
e.vmovdqa(e.ptr[e.rsp + stack_offset_src1], i.src1);
}
e.lea(e.GetNativeParam(0), e.StashXmm(0, src1));
e.CallNativeSafe(reinterpret_cast<void*>(EmulateVectorShr<uint32_t>));
e.vmovaps(i.dest, e.xmm0);
if (i.src2.is_constant) {
e.StashConstantXmm(1, i.src2.constant());
stack_offset_src2 = X64Emitter::kStashOffset + 16;
} else {
e.vmovdqa(e.ptr[e.rsp + stack_offset_src2], i.src2);
}
Xbyak::Label looper;
e.xor_(e.edx, e.edx);
e.L(looper);
e.mov(e.ecx, e.dword[e.rsp + stack_offset_src2 + e.rdx]);
e.shr(e.dword[e.rsp + stack_offset_src1 + e.rdx], e.cl);
e.add(e.edx, 4);
e.cmp(e.edx, 16);
e.jnz(looper);
e.vmovdqa(i.dest, e.byte[e.rsp + stack_offset_src1]);
e.L(end);
}
@@ -1388,7 +1482,8 @@ struct VECTOR_SHA_V128
}
static void EmitInt8(X64Emitter& e, const EmitArgType& i) {
// TODO(benvanik): native version (with shift magic).
unsigned stack_offset_src1 = StackLayout::GUEST_SCRATCH;
unsigned stack_offset_src2 = StackLayout::GUEST_SCRATCH + 16;
if (i.src2.is_constant) {
const auto& shamt = i.src2.constant();
bool all_same = true;
@@ -1399,7 +1494,6 @@ struct VECTOR_SHA_V128
}
}
if (e.IsFeatureEnabled(kX64EmitGFNI)) {
if (all_same) {
// Every count is the same, so we can use gf2p8affineqb.
@@ -1412,8 +1506,7 @@ struct VECTOR_SHA_V128
e.StashConstantXmm(0, vec128q(shift_matrix)), 0);
return;
}
}
else if (all_same) {
} else if (all_same) {
Xmm to_be_shifted = GetInputRegOrConstant(e, i.src1, e.xmm1);
e.vpmovsxbw(e.xmm0, to_be_shifted); //_mm_srai_epi16 / psraw
@@ -1425,14 +1518,41 @@ struct VECTOR_SHA_V128
return;
}
e.lea(e.GetNativeParam(1), e.StashConstantXmm(1, i.src2.constant()));
e.StashConstantXmm(1, i.src2.constant());
stack_offset_src2 = X64Emitter::kStashOffset + 16;
} else {
e.lea(e.GetNativeParam(1), e.StashXmm(1, i.src2));
e.vmovdqa(e.ptr[e.rsp + stack_offset_src2], i.src2);
}
e.lea(e.GetNativeParam(0), e.StashXmm(0, i.src1));
e.CallNativeSafe(reinterpret_cast<void*>(EmulateVectorShr<int8_t>));
e.vmovaps(i.dest, e.xmm0);
if (i.src1.is_constant) {
e.StashConstantXmm(0, i.src1.constant());
stack_offset_src1 = X64Emitter::kStashOffset;
} else {
e.vmovdqa(e.ptr[e.rsp + stack_offset_src1], i.src1);
}
Xbyak::Label looper;
e.xor_(e.edx, e.edx);
e.L(looper);
// movzx is to eliminate any possible dep on previous value of rcx at start
// of loop
e.movzx(e.ecx, e.byte[e.rsp + stack_offset_src2 + e.rdx]);
// maybe using a memory operand as the left side isn't the best idea lol,
// still better than callnativesafe though agners docs have no timing info
// on shx [m], cl so shrug
e.sar(e.byte[e.rsp + stack_offset_src1 + e.rdx], e.cl);
if (e.IsFeatureEnabled(kX64FlagsIndependentVars)) {
e.inc(e.edx);
} else {
e.add(e.edx, 1);
}
e.cmp(e.edx, 16);
e.jnz(looper);
e.vmovdqa(i.dest, e.byte[e.rsp + stack_offset_src1]);
}
static void EmitInt16(X64Emitter& e, const EmitArgType& i) {
@@ -1474,14 +1594,38 @@ struct VECTOR_SHA_V128
// TODO(benvanik): native version (with shift magic).
e.L(emu);
if (i.src2.is_constant) {
e.lea(e.GetNativeParam(1), e.StashConstantXmm(1, i.src2.constant()));
unsigned stack_offset_src1 = StackLayout::GUEST_SCRATCH;
unsigned stack_offset_src2 = StackLayout::GUEST_SCRATCH + 16;
if (i.src1.is_constant) {
e.StashConstantXmm(0, i.src1.constant());
stack_offset_src1 = X64Emitter::kStashOffset;
} else {
e.lea(e.GetNativeParam(1), e.StashXmm(1, i.src2));
e.vmovdqa(e.ptr[e.rsp + stack_offset_src1], i.src1);
}
e.lea(e.GetNativeParam(0), e.StashXmm(0, i.src1));
e.CallNativeSafe(reinterpret_cast<void*>(EmulateVectorShr<int16_t>));
e.vmovaps(i.dest, e.xmm0);
if (i.src2.is_constant) {
e.StashConstantXmm(1, i.src2.constant());
stack_offset_src2 = X64Emitter::kStashOffset + 16;
} else {
e.vmovdqa(e.ptr[e.rsp + stack_offset_src2], i.src2);
}
Xbyak::Label looper;
e.xor_(e.edx, e.edx);
e.L(looper);
e.movzx(e.ecx, e.word[e.rsp + stack_offset_src2 + e.rdx]);
e.sar(e.word[e.rsp + stack_offset_src1 + e.rdx], e.cl);
e.add(e.edx, 2);
e.cmp(e.edx, 16);
e.jnz(looper);
e.vmovdqa(i.dest, e.byte[e.rsp + stack_offset_src1]);
e.L(end);
}
@@ -1508,9 +1652,9 @@ struct VECTOR_SHA_V128
// that happens so we mask.
if (i.src2.is_constant) {
e.LoadConstantXmm(e.xmm0, i.src2.constant());
e.vandps(e.xmm0, e.GetXmmConstPtr(XMMShiftMaskPS));
e.vpand(e.xmm0, e.GetXmmConstPtr(XMMShiftMaskPS));
} else {
e.vandps(e.xmm0, i.src2, e.GetXmmConstPtr(XMMShiftMaskPS));
e.vpand(e.xmm0, i.src2, e.GetXmmConstPtr(XMMShiftMaskPS));
}
e.vpsravd(i.dest, i.src1, e.xmm0);
} else {
@@ -1535,14 +1679,36 @@ struct VECTOR_SHA_V128
// TODO(benvanik): native version.
e.L(emu);
if (i.src2.is_constant) {
e.lea(e.GetNativeParam(1), e.StashConstantXmm(1, i.src2.constant()));
unsigned stack_offset_src1 = StackLayout::GUEST_SCRATCH;
unsigned stack_offset_src2 = StackLayout::GUEST_SCRATCH + 16;
if (i.src1.is_constant) {
e.StashConstantXmm(0, i.src1.constant());
stack_offset_src1 = X64Emitter::kStashOffset;
} else {
e.lea(e.GetNativeParam(1), e.StashXmm(1, i.src2));
e.vmovdqa(e.ptr[e.rsp + stack_offset_src1], i.src1);
}
e.lea(e.GetNativeParam(0), e.StashXmm(0, i.src1));
e.CallNativeSafe(reinterpret_cast<void*>(EmulateVectorShr<int32_t>));
e.vmovaps(i.dest, e.xmm0);
if (i.src2.is_constant) {
e.StashConstantXmm(1, i.src2.constant());
stack_offset_src2 = X64Emitter::kStashOffset + 16;
} else {
e.vmovdqa(e.ptr[e.rsp + stack_offset_src2], i.src2);
}
Xbyak::Label looper;
e.xor_(e.edx, e.edx);
e.L(looper);
e.mov(e.ecx, e.dword[e.rsp + stack_offset_src2 + e.rdx]);
e.sar(e.dword[e.rsp + stack_offset_src1 + e.rdx], e.cl);
e.add(e.edx, 4);
e.cmp(e.edx, 16);
e.jnz(looper);
e.vmovdqa(i.dest, e.byte[e.rsp + stack_offset_src1]);
e.L(end);
}
@@ -1550,26 +1716,6 @@ struct VECTOR_SHA_V128
};
EMITTER_OPCODE_TABLE(OPCODE_VECTOR_SHA, VECTOR_SHA_V128);
// ============================================================================
// OPCODE_VECTOR_ROTATE_LEFT
// ============================================================================
template <typename T, std::enable_if_t<std::is_integral<T>::value, int> = 0>
static __m128i EmulateVectorRotateLeft(void*, __m128i src1, __m128i src2) {
alignas(16) T value[16 / sizeof(T)];
alignas(16) T shamt[16 / sizeof(T)];
// Load SSE registers into a C array.
_mm_store_si128(reinterpret_cast<__m128i*>(value), src1);
_mm_store_si128(reinterpret_cast<__m128i*>(shamt), src2);
for (size_t i = 0; i < (16 / sizeof(T)); ++i) {
value[i] = xe::rotate_left<T>(value[i], shamt[i] & ((sizeof(T) * 8) - 1));
}
// Store result and return it.
return _mm_load_si128(reinterpret_cast<__m128i*>(value));
}
struct VECTOR_ROTATE_LEFT_V128
: Sequence<VECTOR_ROTATE_LEFT_V128,
I<OPCODE_VECTOR_ROTATE_LEFT, V128Op, V128Op, V128Op>> {
@@ -1594,33 +1740,72 @@ struct VECTOR_ROTATE_LEFT_V128
}
} else {
unsigned stack_offset_src1 = StackLayout::GUEST_SCRATCH;
unsigned stack_offset_src2 = StackLayout::GUEST_SCRATCH + 16;
switch (i.instr->flags) {
case INT8_TYPE:
// TODO(benvanik): native version (with shift magic).
if (i.src2.is_constant) {
e.lea(e.GetNativeParam(1),
e.StashConstantXmm(1, i.src2.constant()));
case INT8_TYPE: {
if (i.src1.is_constant) {
e.StashConstantXmm(0, i.src1.constant());
stack_offset_src1 = X64Emitter::kStashOffset;
} else {
e.lea(e.GetNativeParam(1), e.StashXmm(1, i.src2));
e.vmovdqa(e.ptr[e.rsp + stack_offset_src1], i.src1);
}
e.lea(e.GetNativeParam(0), e.StashXmm(0, i.src1));
e.CallNativeSafe(
reinterpret_cast<void*>(EmulateVectorRotateLeft<uint8_t>));
e.vmovaps(i.dest, e.xmm0);
break;
case INT16_TYPE:
// TODO(benvanik): native version (with shift magic).
if (i.src2.is_constant) {
e.lea(e.GetNativeParam(1),
e.StashConstantXmm(1, i.src2.constant()));
e.StashConstantXmm(1, i.src2.constant());
stack_offset_src2 = X64Emitter::kStashOffset + 16;
} else {
e.lea(e.GetNativeParam(1), e.StashXmm(1, i.src2));
e.vmovdqa(e.ptr[e.rsp + stack_offset_src2], i.src2);
}
e.lea(e.GetNativeParam(0), e.StashXmm(0, i.src1));
e.CallNativeSafe(
reinterpret_cast<void*>(EmulateVectorRotateLeft<uint16_t>));
e.vmovaps(i.dest, e.xmm0);
break;
Xbyak::Label rotate_iter;
e.xor_(e.edx, e.edx);
e.L(rotate_iter);
e.movzx(e.ecx, e.byte[e.rsp + stack_offset_src2 + e.rdx]);
e.rol(e.byte[e.rsp + stack_offset_src1 + e.rdx], e.cl);
e.add(e.edx, 1);
e.cmp(e.edx, 16);
e.jnz(rotate_iter);
e.vmovdqa(i.dest, e.byte[e.rsp + stack_offset_src1]);
} break;
case INT16_TYPE: {
if (i.src1.is_constant) {
e.StashConstantXmm(0, i.src1.constant());
stack_offset_src1 = X64Emitter::kStashOffset;
} else {
e.vmovdqa(e.ptr[e.rsp + stack_offset_src1], i.src1);
}
if (i.src2.is_constant) {
e.StashConstantXmm(1, i.src2.constant());
stack_offset_src2 = X64Emitter::kStashOffset + 16;
} else {
e.vmovdqa(e.ptr[e.rsp + stack_offset_src2], i.src2);
}
Xbyak::Label rotate_iter;
e.xor_(e.edx, e.edx);
e.L(rotate_iter);
e.movzx(e.ecx, e.word[e.rsp + stack_offset_src2 + e.rdx]);
e.rol(e.word[e.rsp + stack_offset_src1 + e.rdx], e.cl);
e.add(e.edx, 2);
e.cmp(e.edx, 16);
e.jnz(rotate_iter);
e.vmovdqa(i.dest, e.byte[e.rsp + stack_offset_src1]);
} break;
case INT32_TYPE: {
if (e.IsFeatureEnabled(kX64EmitAVX512Ortho)) {
e.vprolvd(i.dest, i.src1, i.src2);
@@ -1638,23 +1823,40 @@ struct VECTOR_ROTATE_LEFT_V128
}
e.vpsllvd(e.xmm1, i.src1, e.xmm0);
// Shift right (to get low bits):
e.vmovaps(temp, e.GetXmmConstPtr(XMMPI32));
e.vmovdqa(temp, e.GetXmmConstPtr(XMMPI32));
e.vpsubd(temp, e.xmm0);
e.vpsrlvd(i.dest, i.src1, temp);
// Merge:
e.vpor(i.dest, e.xmm1);
} else {
// TODO(benvanik): non-AVX2 native version.
if (i.src2.is_constant) {
e.lea(e.GetNativeParam(1),
e.StashConstantXmm(1, i.src2.constant()));
if (i.src1.is_constant) {
e.StashConstantXmm(0, i.src1.constant());
stack_offset_src1 = X64Emitter::kStashOffset;
} else {
e.lea(e.GetNativeParam(1), e.StashXmm(1, i.src2));
e.vmovdqa(e.ptr[e.rsp + stack_offset_src1], i.src1);
}
e.lea(e.GetNativeParam(0), e.StashXmm(0, i.src1));
e.CallNativeSafe(
reinterpret_cast<void*>(EmulateVectorRotateLeft<uint32_t>));
e.vmovaps(i.dest, e.xmm0);
if (i.src2.is_constant) {
e.StashConstantXmm(1, i.src2.constant());
stack_offset_src2 = X64Emitter::kStashOffset + 16;
} else {
e.vmovdqa(e.ptr[e.rsp + stack_offset_src2], i.src2);
}
Xbyak::Label rotate_iter;
e.xor_(e.edx, e.edx);
e.L(rotate_iter);
e.mov(e.ecx, e.dword[e.rsp + stack_offset_src2 + e.rdx]);
e.rol(e.dword[e.rsp + stack_offset_src1 + e.rdx], e.cl);
e.add(e.edx, 4);
e.cmp(e.edx, 16);
e.jnz(rotate_iter);
e.vmovdqa(i.dest, e.byte[e.rsp + stack_offset_src1]);
}
break;
}
@@ -1667,80 +1869,120 @@ struct VECTOR_ROTATE_LEFT_V128
};
EMITTER_OPCODE_TABLE(OPCODE_VECTOR_ROTATE_LEFT, VECTOR_ROTATE_LEFT_V128);
// ============================================================================
// OPCODE_VECTOR_AVERAGE
// ============================================================================
template <typename T, std::enable_if_t<std::is_integral<T>::value, int> = 0>
static __m128i EmulateVectorAverage(void*, __m128i src1, __m128i src2) {
alignas(16) T src1v[16 / sizeof(T)];
alignas(16) T src2v[16 / sizeof(T)];
alignas(16) T value[16 / sizeof(T)];
// Load SSE registers into a C array.
_mm_store_si128(reinterpret_cast<__m128i*>(src1v), src1);
_mm_store_si128(reinterpret_cast<__m128i*>(src2v), src2);
for (size_t i = 0; i < (16 / sizeof(T)); ++i) {
auto t = (uint64_t(src1v[i]) + uint64_t(src2v[i]) + 1) / 2;
value[i] = T(t);
}
// Store result and return it.
return _mm_load_si128(reinterpret_cast<__m128i*>(value));
}
struct VECTOR_AVERAGE
: Sequence<VECTOR_AVERAGE,
I<OPCODE_VECTOR_AVERAGE, V128Op, V128Op, V128Op>> {
static void Emit(X64Emitter& e, const EmitArgType& i) {
auto i_flags = i.instr->flags;
EmitCommutativeBinaryXmmOp(
e, i,
[&i](X64Emitter& e, const Xmm& dest, const Xmm& src1, const Xmm& src2) {
const TypeName part_type =
static_cast<TypeName>(i.instr->flags & 0xFF);
const uint32_t arithmetic_flags = i.instr->flags >> 8;
[i_flags](X64Emitter& e, const Xmm& dest, const Xmm& src1,
const Xmm& src2) {
const TypeName part_type = static_cast<TypeName>(i_flags & 0xFF);
const uint32_t arithmetic_flags = i_flags >> 8;
bool is_unsigned = !!(arithmetic_flags & ARITHMETIC_UNSIGNED);
unsigned stack_offset_src1 = StackLayout::GUEST_SCRATCH;
unsigned stack_offset_src2 = StackLayout::GUEST_SCRATCH + 16;
switch (part_type) {
case INT8_TYPE:
if (is_unsigned) {
e.vpavgb(dest, src1, src2);
} else {
assert_always();
// todo: avx2 version or version that sign extends to two __m128
e.vmovdqa(e.ptr[e.rsp + stack_offset_src1], src1);
e.vmovdqa(e.ptr[e.rsp + stack_offset_src2], src2);
Xbyak::Label looper;
e.xor_(e.edx, e.edx);
e.L(looper);
e.movsx(e.ecx, e.byte[e.rsp + stack_offset_src2 + e.rdx]);
e.movsx(e.eax, e.byte[e.rsp + stack_offset_src1 + e.rdx]);
e.lea(e.ecx, e.ptr[e.ecx + e.eax + 1]);
e.sar(e.ecx, 1);
e.mov(e.byte[e.rsp + stack_offset_src1 + e.rdx], e.cl);
if (e.IsFeatureEnabled(kX64FlagsIndependentVars)) {
e.inc(e.edx);
} else {
e.add(e.edx, 1);
}
e.cmp(e.edx, 16);
e.jnz(looper);
e.vmovdqa(dest, e.ptr[e.rsp + stack_offset_src1]);
}
break;
case INT16_TYPE:
if (is_unsigned) {
e.vpavgw(dest, src1, src2);
} else {
assert_always();
e.vmovdqa(e.ptr[e.rsp + stack_offset_src1], src1);
e.vmovdqa(e.ptr[e.rsp + stack_offset_src2], src2);
Xbyak::Label looper;
e.xor_(e.edx, e.edx);
e.L(looper);
e.movsx(e.ecx, e.word[e.rsp + stack_offset_src2 + e.rdx]);
e.movsx(e.eax, e.word[e.rsp + stack_offset_src1 + e.rdx]);
e.lea(e.ecx, e.ptr[e.ecx + e.eax + 1]);
e.sar(e.ecx, 1);
e.mov(e.word[e.rsp + stack_offset_src1 + e.rdx], e.cx);
e.add(e.edx, 2);
e.cmp(e.edx, 16);
e.jnz(looper);
e.vmovdqa(dest, e.ptr[e.rsp + stack_offset_src1]);
}
break;
case INT32_TYPE:
case INT32_TYPE: {
// No 32bit averages in AVX.
e.vmovdqa(e.ptr[e.rsp + stack_offset_src1], src1);
e.vmovdqa(e.ptr[e.rsp + stack_offset_src2], src2);
Xbyak::Label looper;
e.xor_(e.edx, e.edx);
e.L(looper);
auto src2_current_ptr =
e.dword[e.rsp + stack_offset_src2 + e.rdx];
auto src1_current_ptr =
e.dword[e.rsp + stack_offset_src1 + e.rdx];
if (is_unsigned) {
if (i.src2.is_constant) {
e.lea(e.GetNativeParam(1),
e.StashConstantXmm(1, i.src2.constant()));
} else {
e.lea(e.GetNativeParam(1), e.StashXmm(1, i.src2));
}
e.lea(e.GetNativeParam(0), e.StashXmm(0, i.src1));
e.CallNativeSafe(
reinterpret_cast<void*>(EmulateVectorAverage<uint32_t>));
e.vmovaps(i.dest, e.xmm0);
// implicit zero-ext
e.mov(e.ecx, src2_current_ptr);
e.mov(e.eax, src1_current_ptr);
} else {
if (i.src2.is_constant) {
e.lea(e.GetNativeParam(1),
e.StashConstantXmm(1, i.src2.constant()));
} else {
e.lea(e.GetNativeParam(1), e.StashXmm(1, i.src2));
}
e.lea(e.GetNativeParam(0), e.StashXmm(0, i.src1));
e.CallNativeSafe(
reinterpret_cast<void*>(EmulateVectorAverage<int32_t>));
e.vmovaps(i.dest, e.xmm0);
e.movsxd(e.rcx, src2_current_ptr);
e.movsxd(e.rax, src1_current_ptr);
}
break;
e.lea(e.rcx, e.ptr[e.rcx + e.rax + 1]);
if (is_unsigned) {
e.shr(e.rcx, 1);
} else {
e.sar(e.rcx, 1);
}
e.mov(e.dword[e.rsp + stack_offset_src1 + e.rdx], e.ecx);
e.add(e.edx, 4);
e.cmp(e.edx, 16);
e.jnz(looper);
e.vmovdqa(dest, e.ptr[e.rsp + stack_offset_src1]);
} break;
default:
assert_unhandled_case(part_type);
break;
@@ -2163,82 +2405,6 @@ struct PERMUTE_V128
};
EMITTER_OPCODE_TABLE(OPCODE_PERMUTE, PERMUTE_I32, PERMUTE_V128);
#define LCPI(name, quad1) const __m128i name = _mm_set1_epi32(quad1)
// xmm0 is precasted to int, but contains float
// chrispy: todo: make available to gpu code
static __m128i xenos_float4_to_float16_x4(__m128i xmm0) {
LCPI(LCPI0_0, 2147483647);
LCPI(LCPI0_1, 1207951360);
LCPI(LCPI0_2, 134217728);
LCPI(LCPI0_3, 3347054592);
LCPI(LCPI0_4, 260038655);
LCPI(LCPI0_5, 32767);
LCPI(LCPI0_6, 4294934528);
__m128i xmm1 = _mm_and_si128(xmm0, LCPI0_0);
__m128i xmm2 = LCPI0_1;
__m128i xmm3 = _mm_add_epi32(xmm0, LCPI0_2);
xmm2 = _mm_cmpgt_epi32(xmm2, xmm1);
xmm3 = _mm_srli_epi32(xmm3, 13);
xmm1 = _mm_add_epi32(xmm1, LCPI0_3);
__m128i xmm4 = _mm_min_epu32(xmm1, LCPI0_4);
xmm1 = _mm_cmpeq_epi32(xmm1, xmm4);
xmm4 = LCPI0_5;
xmm3 = _mm_and_si128(xmm3, xmm4);
xmm1 = _mm_and_si128(xmm1, xmm3);
xmm1 = _mm_castps_si128(_mm_blendv_ps(
_mm_castsi128_ps(xmm4), _mm_castsi128_ps(xmm1), _mm_castsi128_ps(xmm2)));
xmm0 = _mm_srli_epi32(xmm0, 16);
xmm0 = _mm_and_si128(xmm0, LCPI0_6);
xmm0 = _mm_or_si128(xmm1, xmm0);
xmm0 = _mm_packus_epi32(xmm0, _mm_setzero_si128());
return xmm0;
}
// returns floats, uncasted
// chrispy: todo, make this available to gpu code?
static __m128i xenos_halves_to_floats(__m128i xmm0) {
LCPI(LCPI3_0, 0x1f);
LCPI(LCPI3_1, 0x80000000);
LCPI(LCPI3_2, 0x38000000);
LCPI(LCPI3_3, 0x7fe000);
__m128i xmm1, xmm2, xmm3, xmm4;
xmm1 = _mm_cvtepu16_epi32(xmm0);
xmm2 = _mm_srli_epi32(xmm1, 10);
xmm2 = _mm_and_si128(xmm2, LCPI3_0);
xmm0 = _mm_cvtepi16_epi32(xmm0);
xmm0 = _mm_and_si128(xmm0, LCPI3_1);
xmm3 = _mm_setzero_si128();
xmm4 = _mm_slli_epi32(xmm2, 23);
xmm4 = _mm_add_epi32(xmm4, LCPI3_2);
xmm2 = _mm_cmpeq_epi32(xmm2, xmm3);
xmm1 = _mm_slli_epi32(xmm1, 13);
xmm1 = _mm_and_si128(xmm1, LCPI3_3);
xmm3 = _mm_andnot_si128(xmm2, xmm4);
xmm1 = _mm_andnot_si128(xmm2, xmm1);
xmm0 = _mm_or_si128(xmm1, xmm0);
xmm0 = _mm_or_si128(xmm0, xmm3);
return xmm0;
}
#undef LCPI
template <typename Inst>
static void emit_fast_f16_unpack(X64Emitter& e, const Inst& i,
XmmConst initial_shuffle) {

104
src/xenia/cpu/backend/x64/x64_sequences.cc
View File

File diff suppressed because one or more lines are too long