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

remove useless tag field from hir::Value

Browse Source 
pack local_slot and constant in hir::Value
Instead of loading membase at the start of every function, just load it in HostToGuestThunk
vzeroupper in GuestToHostThunk before calling host function, and in HostToGuestThunk after calling function to prevent AVX dirty state slowdowns. In the future, check if CPU implements AVX as 128x2 and skip if so (https://john-h-k.github.io/VexTransitionPenalties.html)
Remove useless save/restore of ctx pointer, nothing modifies it and it prevents cpus from doing cross-function memory renaming (https://www.agner.org/forum/viewtopic.php?t=41). Could not remove the space on stack because of alignment issues, instead turned it into GUEST_SCRATCH64 which is a temporary that sequences may use
Reorder OpcodeInfo so that name is at offset 0, remove name and add GetOpcodeName function (name is only used for debug code, we are seperating frequently accessed data and rarely accessed data)
Add VECTOR_DENORMFLUSH opcode for handling output to DOT_PRODUCT and other opcodes that implicitly force denormal inputs/outputs to zero, will eventually use for implementing NJM
Rewrite sequences for LOAD_VECTOR_SHL/SHR. The mask with 0xf in it was pointless as all InstrEmit_ functions that create the load shift instructions do that in HIR. The tables are only used for nonzero constant inputs now, which are probably pretty rare. Instead of doing a shift and lookup, a base value is used for both in the constant table and adding/subtracting of the input is done
Reuse result of LoadVectorShl/Shr in InstrEmit_stvlx_, InstrEmit_stvrx_. We were previously calculating it twice which was contributing to the final sequences' fatness. Use OPCODE_SELECT instead of the sequence of or, andnot, and that it was using for merging
Add the proper unconditional denormal input flushing behavior to vfmadd, add it also to vfmsub (making the assumption it has the same behavior)
Remove constant propagation for DOT_PRODUCT_3/4
DOT_PRODUCT_3/4 now returns a vector with all four elements set to the result. (what we were doing before, truncating to float32 and then splatting didnt make any sense)
Add much more correct versions of DOT_PRODUCT_3/4, matching the Xb360's  to 1 bit. Still needs work to be a perfect emulation.
Add constant folding for OPCODE_SELECT, OPCODE_INSERT, OPCODE_PERMUTE, OPCODE_SWIZZLE
Remove constant folding for DOT_PRODUCT
Removed the multibyte nop code I committed earlier, it doesnt help us much because nops are only used for debug stuff and its ugly and wouldnt survive in a pr to main
Check for AVX512BMI, use vpermb to shuffle if supported
This commit is contained in:
chss95cs@gmail.com
2022-07-16 10:25:04 -07:00
parent 5f11c5d3b4
commit 6a612b4d34
17 changed files with 574 additions and 231 deletions
Download Patch File Download Diff File Expand all files Collapse all files

10
src/xenia/cpu/backend/x64/x64_backend.cc
View File

@@ -446,10 +446,11 @@ HostToGuestThunk X64ThunkEmitter::EmitHostToGuestThunk() {
EmitSaveNonvolatileRegs();
mov(rax, rcx);
mov(rsi, rdx); // context
mov(rcx, r8); // return address
mov(rsi, rdx); // context
mov(rdi, ptr[rdx + offsetof(ppc::PPCContext, virtual_membase)]); // membase
mov(rcx, r8); // return address
call(rax);
vzeroupper();
EmitLoadNonvolatileRegs();
code_offsets.epilog = getSize();
@@ -500,7 +501,8 @@ GuestToHostThunk X64ThunkEmitter::EmitGuestToHostThunk() {
code_offsets.prolog_stack_alloc = getSize();
code_offsets.body = getSize();
// chrispy: added this for proper vmsum impl, avx2 bitshifts
vzeroupper();
// Save off volatile registers.
EmitSaveVolatileRegs();

101
src/xenia/cpu/backend/x64/x64_emitter.cc
View File

@@ -101,13 +101,11 @@ X64Emitter::X64Emitter(X64Backend* backend, XbyakAllocator* allocator)
TEST_EMIT_FEATURE(kX64EmitAVX512VL, Xbyak::util::Cpu::tAVX512VL);
TEST_EMIT_FEATURE(kX64EmitAVX512BW, Xbyak::util::Cpu::tAVX512BW);
TEST_EMIT_FEATURE(kX64EmitAVX512DQ, Xbyak::util::Cpu::tAVX512DQ);
TEST_EMIT_FEATURE(kX64EmitAVX512VBMI, Xbyak::util::Cpu::tAVX512VBMI);
#undef TEST_EMIT_FEATURE
/*
fix for xbyak bug/omission, amd cpus are never checked for lzcnt. fixed in latest version of xbyak
fix for xbyak bug/omission, amd cpus are never checked for lzcnt. fixed in
latest version of xbyak
*/
unsigned int data[4];
Xbyak::util::Cpu::getCpuid(0x80000001, data);
@@ -117,21 +115,19 @@ X64Emitter::X64Emitter(X64Backend* backend, XbyakAllocator* allocator)
}
}
if (cpu_.has(Xbyak::util::Cpu::tAMD)) {
bool is_zennish = cpu_.displayFamily >= 0x17;
bool is_zennish = cpu_.displayFamily >= 0x17;
if (is_zennish) {
feature_flags_ |= kX64FastJrcx;
if (is_zennish) {
feature_flags_ |= kX64FastJrcx;
if (cpu_.displayFamily > 0x17) {
feature_flags_ |= kX64FastLoop;
if (cpu_.displayFamily > 0x17) {
feature_flags_ |= kX64FastLoop;
} else if (cpu_.displayFamily == 0x17 && cpu_.displayModel >= 0x31) {
feature_flags_ |= kX64FastLoop;
} // todo:figure out at model zen+ became zen2, this is just the model
// for my cpu, which is ripper90
}
} else if (cpu_.displayFamily == 0x17 && cpu_.displayModel >= 0x31) {
feature_flags_ |= kX64FastLoop;
} // todo:figure out at model zen+ became zen2, this is just the model
// for my cpu, which is ripper90
}
}
}
@@ -263,7 +259,10 @@ bool X64Emitter::Emit(HIRBuilder* builder, EmitFunctionInfo& func_info) {
code_offsets.prolog_stack_alloc = getSize();
code_offsets.body = getSize();
/*
* chrispy: removed this, it serves no purpose
mov(qword[rsp + StackLayout::GUEST_CTX_HOME], GetContextReg());
*/
mov(qword[rsp + StackLayout::GUEST_RET_ADDR], rcx);
mov(qword[rsp + StackLayout::GUEST_CALL_RET_ADDR], 0);
@@ -296,9 +295,11 @@ bool X64Emitter::Emit(HIRBuilder* builder, EmitFunctionInfo& func_info) {
}
// Load membase.
mov(GetMembaseReg(),
/*
* chrispy: removed this, as long as we load it in HostToGuestThunk we can
count on no other code modifying it. mov(GetMembaseReg(),
qword[GetContextReg() + offsetof(ppc::PPCContext, virtual_membase)]);
*/
// Body.
auto block = builder->first_block();
while (block) {
@@ -318,7 +319,7 @@ bool X64Emitter::Emit(HIRBuilder* builder, EmitFunctionInfo& func_info) {
// NOTE: If you encounter this after adding a new instruction, do a full
// rebuild!
assert_always();
XELOGE("Unable to process HIR opcode {}", instr->opcode->name);
XELOGE("Unable to process HIR opcode {}", GetOpcodeName(instr->opcode));
break;
}
instr = new_tail;
@@ -331,8 +332,10 @@ bool X64Emitter::Emit(HIRBuilder* builder, EmitFunctionInfo& func_info) {
L(epilog_label);
epilog_label_ = nullptr;
EmitTraceUserCallReturn();
/*
* chrispy: removed this, it serves no purpose
mov(GetContextReg(), qword[rsp + StackLayout::GUEST_CTX_HOME]);
*/
code_offsets.epilog = getSize();
add(rsp, (uint32_t)stack_size);
@@ -342,7 +345,6 @@ bool X64Emitter::Emit(HIRBuilder* builder, EmitFunctionInfo& func_info) {
if (cvars::emit_source_annotations) {
nop(5);
}
assert_zero(code_offsets.prolog);
@@ -676,37 +678,9 @@ Xbyak::Reg64 X64Emitter::GetNativeParam(uint32_t param) {
Xbyak::Reg64 X64Emitter::GetContextReg() { return rsi; }
Xbyak::Reg64 X64Emitter::GetMembaseReg() { return rdi; }
void X64Emitter::ReloadContext() {
mov(GetContextReg(), qword[rsp + StackLayout::GUEST_CTX_HOME]);
}
void X64Emitter::ReloadMembase() {
mov(GetMembaseReg(), qword[GetContextReg() + 8]); // membase
}
#define __NH_CONCAT(x, y) x##y
#define _MH_CONCAT(cb, ...) cb (__VA_ARGS__)
#define mh_concat2_m(x, y) __NH_CONCAT(x, y)
#define DECLNOP(n, ...) \
static constexpr unsigned char mh_concat2_m(nop_, n)[] = {__VA_ARGS__}
DECLNOP(1, 0x90);
DECLNOP(2, 0x66, 0x90);
DECLNOP(3, 0x0F, 0x1F, 0x00);
DECLNOP(4, 0x0F, 0x1F, 0x40, 0x00);
DECLNOP(5, 0x0F, 0x1F, 0x44, 0x00, 0x00);
DECLNOP(6, 0x66, 0x0F, 0x1F, 0x44, 0x00, 0x00);
DECLNOP(7, 0x0F, 0x1F, 0x80, 0x00, 0x00, 0x00, 0x00);
DECLNOP(8, 0x0F, 0x1F, 0x84, 00, 00, 00, 00, 00);
DECLNOP(9, 0x66, 0x0F, 0x1F, 0x84, 00, 00, 00, 00, 00);
static constexpr const unsigned char* const g_noptable[] = {
&nop_1[0], &nop_1[0], &nop_2[0], &nop_3[0], &nop_4[0],
&nop_5[0], &nop_6[0], &nop_7[0], &nop_8[0], &nop_9[0]};
static constexpr unsigned LENGTHOF_NOPTABLE =
sizeof(g_noptable) / sizeof(g_noptable[0]);
// Len Assembly Byte Sequence
// ============================================================================
@@ -720,17 +694,8 @@ static constexpr unsigned LENGTHOF_NOPTABLE =
// 8b NOP DWORD ptr [EAX + EAX*1 + 00000000H] 0F 1F 84 00 00 00 00 00H
// 9b 66 NOP DWORD ptr [EAX + EAX*1 + 00000000H] 66 0F 1F 84 00 00 00 00 00H
void X64Emitter::nop(size_t length) {
while (length != 0) {
unsigned patchsize = length % LENGTHOF_NOPTABLE;
// patch_memory(locptr, size, (char*)g_noptable[patchsize]);
for (unsigned i = 0; i < patchsize; ++i) {
db(g_noptable[patchsize][i]);
}
//locptr += patchsize;
length -= patchsize;
for (size_t i = 0; i < length; ++i) {
db(0x90);
}
}
@@ -912,8 +877,17 @@ static const vec128_t xmm_consts[] = {
0x80, 0x80, 0x80, 0x80),
/*XMMShortsToBytes*/
v128_setr_bytes(0, 2, 4, 6, 8, 10, 12, 14, 0x80, 0x80, 0x80, 0x80, 0x80,
0x80, 0x80, 0x80)
};
0x80, 0x80, 0x80),
/*XMMLVSLTableBase*/
vec128b(0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15),
/*XMMLVSRTableBase*/
vec128b(16, 17, 18, 19, 20, 21, 22, 23, 24, 25, 26, 27, 28, 29, 30, 31),
/* XMMSingleDenormalMask */
vec128i(0x7f800000),
/* XMMThreeFloatMask */
vec128i(~0U, ~0U, ~0U, 0U),
/*XMMXenosF16ExtRangeStart*/
vec128f(65504)};
void* X64Emitter::FindByteConstantOffset(unsigned bytevalue) {
for (auto& vec : xmm_consts) {
@@ -1013,7 +987,6 @@ void X64Emitter::LoadConstantXmm(Xbyak::Xmm dest, const vec128_t& v) {
// 1111...
vpcmpeqb(dest, dest);
} else {
for (size_t i = 0; i < (kConstDataSize / sizeof(vec128_t)); ++i) {
if (xmm_consts[i] == v) {
vmovapd(dest, GetXmmConstPtr((XmmConst)i));

10
src/xenia/cpu/backend/x64/x64_emitter.h
View File

@@ -118,7 +118,12 @@ enum XmmConst {
XMM2To32,
XMMFloatInf,
XMMIntsToBytes,
XMMShortsToBytes
XMMShortsToBytes,
XMMLVSLTableBase,
XMMLVSRTableBase,
XMMSingleDenormalMask,
XMMThreeFloatMask, //for clearing the fourth float prior to DOT_PRODUCT_3
XMMXenosF16ExtRangeStart
};
// Unfortunately due to the design of xbyak we have to pass this to the ctor.
@@ -147,6 +152,7 @@ enum X64EmitterFeatureFlags {
kX64EmitAVX512Ortho64 = kX64EmitAVX512Ortho | kX64EmitAVX512DQ,
kX64FastJrcx = 1 << 12, //jrcxz is as fast as any other jump ( >= Zen1)
kX64FastLoop = 1 << 13, //loop/loope/loopne is as fast as any other jump ( >= Zen2)
kX64EmitAVX512VBMI = 1 << 14
};
class ResolvableGuestCall {
public:
@@ -225,7 +231,7 @@ class X64Emitter : public Xbyak::CodeGenerator {
Xbyak::Reg64 GetContextReg();
Xbyak::Reg64 GetMembaseReg();
void ReloadContext();
void ReloadMembase();
void nop(size_t length = 1);

74
src/xenia/cpu/backend/x64/x64_seq_vector.cc
View File

@@ -127,6 +127,26 @@ struct VECTOR_CONVERT_F2I
};
EMITTER_OPCODE_TABLE(OPCODE_VECTOR_CONVERT_F2I, VECTOR_CONVERT_F2I);
struct VECTOR_DENORMFLUSH
: Sequence<VECTOR_DENORMFLUSH,
I<OPCODE_VECTOR_DENORMFLUSH, V128Op, V128Op>> {
static void Emit(X64Emitter& e, const EmitArgType& i) {
e.vxorps(e.xmm1, e.xmm1, e.xmm1); // 0.25 P0123
e.vandps(e.xmm0, i.src1,
e.GetXmmConstPtr(XMMSingleDenormalMask)); // 0.25 P0123
e.vcmpneqps(e.xmm2, e.xmm0, e.xmm1); // 0.5 P01
e.vandps(e.xmm1, i.src1,
e.GetXmmConstPtr(XMMSignMaskF32)); // 0.5 P0123 take signs, zeros
// must keep their signs
e.vandps(e.xmm0, i.src1, e.xmm2); // P0123
e.vorps(i.dest, e.xmm0, e.xmm1); // P0123 make sure zeros keep signs
// if it does not equal zero, we stay
}
};
EMITTER_OPCODE_TABLE(OPCODE_VECTOR_DENORMFLUSH, VECTOR_DENORMFLUSH);
// ============================================================================
// OPCODE_LOAD_VECTOR_SHL
// ============================================================================
@@ -154,15 +174,20 @@ struct LOAD_VECTOR_SHL_I8
if (i.src1.is_constant) {
auto sh = i.src1.constant();
assert_true(sh < xe::countof(lvsl_table));
e.mov(e.rax, (uintptr_t)&lvsl_table[sh]);
e.vmovaps(i.dest, e.ptr[e.rax]);
if (sh == 0) {
e.vmovdqa(i.dest, e.GetXmmConstPtr(XMMLVSLTableBase));
} else {
// this is probably extremely rare
e.LoadConstantXmm(i.dest, lvsl_table[sh]);
}
} else {
// TODO(benvanik): find a cheaper way of doing this.
e.movzx(e.rdx, i.src1);
e.and_(e.dx, 0xF);
e.shl(e.dx, 4);
e.mov(e.rax, (uintptr_t)lvsl_table);
e.vmovaps(i.dest, e.ptr[e.rax + e.rdx]);
// chrispy: removed mask, ppc_emit_altivec already pre-ands it.
e.vmovd(e.xmm0, i.src1.reg().cvt32());
// broadcast byte
// dont use broadcastb with avx2, its slower than shuf
e.vpshufb(e.xmm0, e.xmm0, e.GetXmmConstPtr(XMMZero));
e.vpaddb(i.dest, e.xmm0, e.GetXmmConstPtr(XMMLVSLTableBase));
}
}
};
@@ -195,15 +220,23 @@ struct LOAD_VECTOR_SHR_I8
if (i.src1.is_constant) {
auto sh = i.src1.constant();
assert_true(sh < xe::countof(lvsr_table));
e.mov(e.rax, (uintptr_t)&lvsr_table[sh]);
e.vmovaps(i.dest, e.ptr[e.rax]);
if (sh == 0) {
e.vmovdqa(i.dest, e.GetXmmConstPtr(XMMLVSRTableBase));
} else {
e.LoadConstantXmm(i.dest, lvsr_table[sh]);
}
} else {
// TODO(benvanik): find a cheaper way of doing this.
e.movzx(e.rdx, i.src1);
e.and_(e.dx, 0xF);
e.shl(e.dx, 4);
e.mov(e.rax, (uintptr_t)lvsr_table);
e.vmovaps(i.dest, e.ptr[e.rax + e.rdx]);
// chrispy: removed mask, ppc_emit_altivec already pre-ands it. removed
// lookup as well, compute from LVSR base instead
e.vmovd(e.xmm0, i.src1.reg().cvt32());
e.vmovdqa(e.xmm1, e.GetXmmConstPtr(XMMLVSRTableBase));
// broadcast byte
// dont use broadcastb with avx2, its slower than shuf
e.vpshufb(e.xmm0, e.xmm0, e.GetXmmConstPtr(XMMZero));
e.vpsubb(i.dest, e.xmm1, e.xmm0);
}
}
};
@@ -728,7 +761,7 @@ struct VECTOR_SHL_V128
}
}
static void EmitInt8(X64Emitter& e, const EmitArgType& i) {
static void EmitInt8(X64Emitter& e, const EmitArgType& i) {
// TODO(benvanik): native version (with shift magic).
if (e.IsFeatureEnabled(kX64EmitAVX2)) {
@@ -1793,6 +1826,14 @@ struct PERMUTE_I32
}
}
};
//todo: use this on const src1
static vec128_t FixupConstantShuf8(vec128_t input) {
for (uint32_t i = 0; i < 16; ++i) {
input.u8[i] ^= 0x03;
input.u8[i] &= 0x1F;
}
return input;
}
struct PERMUTE_V128
: Sequence<PERMUTE_V128,
I<OPCODE_PERMUTE, V128Op, V128Op, V128Op, V128Op>> {
@@ -1855,7 +1896,8 @@ struct PERMUTE_V128
} else {
e.vpshufb(src3_shuf, i.src3, e.xmm2);
}
// Build a mask with values in src2 having 0 and values in src3 having 1.
// Build a mask with values in src2 having 0 and values in src3
// having 1.
e.vpcmpgtb(i.dest, e.xmm2, e.GetXmmConstPtr(XMMPermuteControl15));
e.vpblendvb(i.dest, src2_shuf, src3_shuf, i.dest);
}

220
src/xenia/cpu/backend/x64/x64_sequences.cc
View File

@@ -35,11 +35,14 @@
#include "xenia/cpu/backend/x64/x64_emitter.h"
#include "xenia/cpu/backend/x64/x64_op.h"
#include "xenia/cpu/backend/x64/x64_tracers.h"
// needed for stmxcsr
#include "xenia/cpu/backend/x64/x64_stack_layout.h"
#include "xenia/cpu/hir/hir_builder.h"
#include "xenia/cpu/processor.h"
DEFINE_bool(use_fast_dot_product, false,
"Experimental optimization, much shorter sequence on dot products, treating inf as overflow instead of using mcxsr"
"Experimental optimization, much shorter sequence on dot products, "
"treating inf as overflow instead of using mcxsr"
"four insn dotprod",
"CPU");
namespace xe {
@@ -1996,8 +1999,8 @@ struct DIV_V128 : Sequence<DIV_V128, I<OPCODE_DIV, V128Op, V128Op, V128Op>> {
assert_true(!i.instr->flags);
EmitAssociativeBinaryXmmOp(e, i,
[](X64Emitter& e, Xmm dest, Xmm src1, Xmm src2) {
// e.vrcpps(e.xmm0, src2);
//e.vmulps(dest, src1, e.xmm0);
// e.vrcpps(e.xmm0, src2);
// e.vmulps(dest, src1, e.xmm0);
e.vdivps(dest, src1, src2);
});
}
@@ -2607,68 +2610,84 @@ struct LOG2_V128 : Sequence<LOG2_V128, I<OPCODE_LOG2, V128Op, V128Op>> {
};
EMITTER_OPCODE_TABLE(OPCODE_LOG2, LOG2_F32, LOG2_F64, LOG2_V128);
struct DOT_PRODUCT_V128 {
static void Emit(X64Emitter& e, Xmm dest, Xmm src1, Xmm src2, uint8_t imm) {
if (cvars::use_fast_dot_product) {
e.vdpps(dest, src1, src2, imm);
e.vandps(e.xmm0, dest, e.GetXmmConstPtr(XMMAbsMaskPS));
e.vcmpgeps(e.xmm0, e.xmm0, e.GetXmmConstPtr(XMMFloatInf));
e.vblendvps(dest, dest, e.GetXmmConstPtr(XMMQNaN), e.xmm0);
} else {
// TODO(benvanik): apparently this is very slow
// - find alternative?
Xbyak::Label end;
e.inLocalLabel();
// Grab space to put MXCSR.
// TODO(gibbed): stick this in TLS or
// something?
e.sub(e.rsp, 8);
// Grab MXCSR and mask off the overflow flag,
// because it's sticky.
e.vstmxcsr(e.dword[e.rsp]);
e.mov(e.eax, e.dword[e.rsp]);
e.and_(e.eax, uint32_t(~8));
e.mov(e.dword[e.rsp], e.eax);
e.vldmxcsr(e.dword[e.rsp]);
// Hey we can do the dot product now.
e.vdpps(dest, src1, src2, imm);
// Load MXCSR...
e.vstmxcsr(e.dword[e.rsp]);
// ..free our temporary space and get MXCSR at
// the same time
e.pop(e.rax);
// Did we overflow?
e.test(e.al, 8);
e.jz(end);
// Infinity? HA! Give NAN.
e.vmovdqa(dest, e.GetXmmConstPtr(XMMQNaN));
e.L(end);
e.outLocalLabel();
}
}
};
// ============================================================================
// OPCODE_DOT_PRODUCT_3
// ============================================================================
struct DOT_PRODUCT_3_V128
: Sequence<DOT_PRODUCT_3_V128,
I<OPCODE_DOT_PRODUCT_3, F32Op, V128Op, V128Op>> {
I<OPCODE_DOT_PRODUCT_3, V128Op, V128Op, V128Op>> {
static void Emit(X64Emitter& e, const EmitArgType& i) {
// https://msdn.microsoft.com/en-us/library/bb514054(v=vs.90).aspx
EmitCommutativeBinaryXmmOp(
e, i, [](X64Emitter& e, Xmm dest, Xmm src1, Xmm src2) {
DOT_PRODUCT_V128::Emit(e, dest, src1, src2, 0b01110001);
});
// todo: add fast_dot_product path that just checks for infinity instead of
// using mxcsr
auto mxcsr_storage = e.dword[e.rsp + StackLayout::GUEST_SCRATCH64];
// this is going to hurt a bit...
/*
this implementation is accurate, it matches the results of xb360 vmsum3
except that vmsum3 is often off by 1 bit, but its extremely slow. it is a
long, unbroken chain of dependencies, and the three uses of mxcsr all cost
about 15-20 cycles at the very least on amd zen processors. on older amd the
figures agner has are pretty horrible. it looks like its just as bad on
modern intel cpus also up until just recently. perhaps a better way of
detecting overflow would be to just compare with inf. todo: test whether cmp
with inf can replace
*/
e.vstmxcsr(mxcsr_storage);
e.mov(e.eax, 8);
auto src1v = e.xmm0;
auto src2v = e.xmm1;
if (i.src1.is_constant) {
src1v = e.xmm0;
e.LoadConstantXmm(src1v, i.src1.constant());
} else {
src1v = i.src1.reg();
}
if (i.src2.is_constant) {
src2v = e.xmm1;
e.LoadConstantXmm(src2v, i.src2.constant());
} else {
src2v = i.src2.reg();
}
e.not_(e.eax);
// todo: maybe the top element should be cleared by the InstrEmit_ function
// so that in the future this could be optimized away if the top is known to
// be zero. Right now im not sure that happens often though and its
// currently not worth it also, maybe pre-and if constant
e.vandps(e.xmm3, src1v, e.GetXmmConstPtr(XMMThreeFloatMask));
e.vandps(e.xmm2, src2v, e.GetXmmConstPtr(XMMThreeFloatMask));
e.and_(mxcsr_storage, e.eax);
e.vldmxcsr(mxcsr_storage); // overflow flag is cleared, now we're good to
// go
e.vcvtps2pd(e.ymm0, e.xmm3);
e.vcvtps2pd(e.ymm1, e.xmm2);
/*
ymm0 = src1 as doubles, ele 3 cleared
ymm1 = src2 as doubles, ele 3 cleared
*/
e.vmulpd(e.ymm3, e.ymm0, e.ymm1);
e.vextractf128(e.xmm2, e.ymm3, 1);
e.vunpckhpd(e.xmm0, e.xmm3, e.xmm3); // get element [1] in xmm3
e.vaddsd(e.xmm3, e.xmm3, e.xmm2);
e.not_(e.eax);
e.vaddsd(e.xmm2, e.xmm3, e.xmm0);
e.vcvtsd2ss(e.xmm1, e.xmm2);
// this is awful
e.vstmxcsr(mxcsr_storage);
e.test(mxcsr_storage, e.eax);
Xbyak::Label ret_qnan;
Xbyak::Label done;
e.jnz(ret_qnan);
// e.vshufps(i.dest, e.xmm1,e.xmm1, 0); // broadcast
e.vbroadcastss(i.dest, e.xmm1);
e.jmp(done);
e.L(ret_qnan);
e.vmovaps(i.dest, e.GetXmmConstPtr(XMMQNaN));
e.L(done);
}
};
EMITTER_OPCODE_TABLE(OPCODE_DOT_PRODUCT_3, DOT_PRODUCT_3_V128);
@@ -2678,13 +2697,81 @@ EMITTER_OPCODE_TABLE(OPCODE_DOT_PRODUCT_3, DOT_PRODUCT_3_V128);
// ============================================================================
struct DOT_PRODUCT_4_V128
: Sequence<DOT_PRODUCT_4_V128,
I<OPCODE_DOT_PRODUCT_4, F32Op, V128Op, V128Op>> {
I<OPCODE_DOT_PRODUCT_4, V128Op, V128Op, V128Op>> {
static void Emit(X64Emitter& e, const EmitArgType& i) {
// https://msdn.microsoft.com/en-us/library/bb514054(v=vs.90).aspx
EmitCommutativeBinaryXmmOp(
e, i, [](X64Emitter& e, Xmm dest, Xmm src1, Xmm src2) {
DOT_PRODUCT_V128::Emit(e, dest, src1, src2, 0b11110001);
});
// todo: add fast_dot_product path that just checks for infinity instead of
// using mxcsr
auto mxcsr_storage = e.dword[e.rsp + StackLayout::GUEST_SCRATCH64];
e.vstmxcsr(mxcsr_storage);
e.mov(e.eax, 8);
auto src1v = e.xmm3;
auto src2v = e.xmm2;
if (i.src1.is_constant) {
src1v = e.xmm3;
e.LoadConstantXmm(src1v, i.src1.constant());
} else {
src1v = i.src1.reg();
}
if (i.src2.is_constant) {
src2v = e.xmm2;
e.LoadConstantXmm(src2v, i.src2.constant());
} else {
src2v = i.src2.reg();
}
e.not_(e.eax);
e.and_(mxcsr_storage, e.eax);
e.vldmxcsr(mxcsr_storage);
e.vcvtps2pd(e.ymm0, src1v);
e.vcvtps2pd(e.ymm1, src2v);
/*
e.vandps(e.xmm3, src1v, e.GetXmmConstPtr(XMMThreeFloatMask));
e.vandps(e.xmm2, src2v, e.GetXmmConstPtr(XMMThreeFloatMask));
e.and_(mxcsr_storage, e.eax);
e.vldmxcsr(mxcsr_storage); // overflow flag is cleared, now we're good to
// go
e.vcvtps2pd(e.ymm0, e.xmm3);
e.vcvtps2pd(e.ymm1, e.xmm2);
e.vmulpd(e.ymm5, e.ymm0, e.ymm1);
e.vextractf128(e.xmm4, e.ymm5, 1);
e.vunpckhpd(e.xmm3, e.xmm5, e.xmm5); // get element [1] in xmm3
e.vaddsd(e.xmm5, e.xmm5, e.xmm4);
e.not_(e.eax);
e.vaddsd(e.xmm2, e.xmm5, e.xmm3);
e.vcvtsd2ss(e.xmm1, e.xmm2);
*/
e.vmulpd(e.ymm3, e.ymm0, e.ymm1);
e.vextractf128(e.xmm2, e.ymm3, 1);
e.vaddpd(e.xmm3, e.xmm3, e.xmm2);
e.vunpckhpd(e.xmm0, e.xmm3, e.xmm3);
e.not_(e.eax);
e.vaddsd(e.xmm2, e.xmm3, e.xmm0);
e.vcvtsd2ss(e.xmm1, e.xmm2);
e.vstmxcsr(mxcsr_storage);
e.test(mxcsr_storage, e.eax);
Xbyak::Label ret_qnan;
Xbyak::Label done;
e.jnz(ret_qnan); // reorder these jmps later, just want to get this fix in
// e.vshufps(i.dest, e.xmm1, e.xmm1, 0);
e.vbroadcastss(i.dest, e.xmm1);
e.jmp(done);
e.L(ret_qnan);
e.vmovaps(i.dest, e.GetXmmConstPtr(XMMQNaN));
e.L(done);
// e.DebugBreak();
}
};
EMITTER_OPCODE_TABLE(OPCODE_DOT_PRODUCT_4, DOT_PRODUCT_4_V128);
@@ -2759,7 +2846,6 @@ struct AND_I64 : Sequence<AND_I64, I<OPCODE_AND, I64Op, I64Op, I64Op>> {
};
struct AND_V128 : Sequence<AND_V128, I<OPCODE_AND, V128Op, V128Op, V128Op>> {
static void Emit(X64Emitter& e, const EmitArgType& i) {
EmitCommutativeBinaryXmmOp(e, i,
[](X64Emitter& e, Xmm dest, Xmm src1, Xmm src2) {
e.vpand(dest, src1, src2);
@@ -3419,7 +3505,7 @@ bool SelectSequence(X64Emitter* e, const Instr* i, const Instr** new_tail) {
return true;
}
}
XELOGE("No sequence match for variant {}", i->opcode->name);
XELOGE("No sequence match for variant {}", GetOpcodeName(i->opcode));
return false;
}

3
src/xenia/cpu/backend/x64/x64_stack_layout.h
View File

@@ -122,7 +122,8 @@ class StackLayout {
*
*/
static const size_t GUEST_STACK_SIZE = 104;
static const size_t GUEST_CTX_HOME = 80;
//was GUEST_CTX_HOME, can't remove because that'd throw stack alignment off. instead, can be used as a temporary in sequences
static const size_t GUEST_SCRATCH64 = 80;
static const size_t GUEST_RET_ADDR = 88;
static const size_t GUEST_CALL_RET_ADDR = 96;
};