Add OPCODE_NEGATED_MUL_ADD/OPCODE_NEGATED_MUL_SUB

Proper handling of nans for VMX max/min on x64 (minps/maxps has special behavior depending on the operand order that vmx does not have for vminfp/vmaxfp) Add extremely unintrusive guest code profiler utilizing KUSER_SHARED systemtime. This profiler is disabled on platforms other than windows, and on windows is disabled by default by a cvar Repurpose GUEST_SCRATCH64 stack offset to instead be for storing guest function profile times, define GUEST_SCRATCH as 0 instead, since thats already meant to be a scratch area Fix xenia silently closing on config errors/other fatal errors by setting has_console_attached_'s default to false Add alternative code path for guest clock that uses kusershared systemtime instead of QueryPerformanceCounter. This is way faster and I have tested it and found it to be working, but i have disabled it because i do not know how well it works on wine or on processors other than mine Significantly reduce log spam by setting XELOGAPU and XELOGGPU to be LogLevel::Debug Changed some LOGI to LOGD in places to reduce log spam Mark VdSwap as kHighFrequency, it was spamming up logs Make logging calls less intrusive for the caller by forcing the test of log level inline and moving the format/AppendLogLine stuff to an outlined cold function Add swcache namespace for software cache operations like prefetches, streaming stores and streaming loads. Add XE_MSVC_REORDER_BARRIER for preventing msvc from propagating a value too close to its store or from its load Add xe_unlikely_mutex for locks we know have very little contention add XE_HOST_CACHE_LINE_SIZE and XE_RESTRICT to platform.h Microoptimization: Changed most uses of size_t to ring_size_t in RingBuffer, this reduces the size of the inlined ringbuffer operations slightly by eliminating rex prefixes, depending on register allocation Add BeginPrefetchedRead to ringbuffer, which prefetches the second range if there is one according to the provided PrefetchTag added inline_loadclock cvar, which will directly use the value of the guest clock from clock.cc in jitted guest code. off by default change uses of GUEST_SCRATCH64 to GUEST_SCRATCH Add fast vectorized xenos_half_to_float/xenos_float_to_half (currently resides in x64_seq_vector, move to gpu code maybe at some point) Add fast x64 codegen for PackFloat16_4/UnpackFloat16_4. Same code can be used for Float16_2 in future commit. This should speed up some games that use these functions heavily Remove cvar for toggling old float16 behavior Add VRSAVE register, support mfspr/mtspr vrsave Add cvar for toggling off codegen for trap instructions and set it to true by default. Add specialized methods to CommandProcessor: WriteRegistersFromMem, WriteRegisterRangeFromRing, and WriteOneRegisterFromRing. These reduce the overall cost of WriteRegister Use a fixed size vmem vector for upload ranges, realloc/memsetting on resize in the inner loop of requestranges was showing up on the profiler (the search in requestranges itself needs work) Rename fixed_vmem_vector to better fit xenia's naming convention Only log unknown register writes in WriteRegister if DEBUG :/. We're stuck on MSVC with c++17 so we have no way of influencing the branch ordering for that function without profile guided optimization Remove binding stride assert in shader_translator.cc, triangle told me its leftover ogl stuff Mark xe::FatalError as noreturn If a controller is not connected, delay by 1.1 seconds before checking if it has been reconnected. Asking Xinput about a controller slot that is unused is extremely slow, and XinputGetState/SetState were taking up an enormous amount of time in profiles. this may have caused a bit of input lag Protect accesses to input_system with a lock Add proper handling for user_index>= 4 in XamInputGetState/SetState, properly return zeroed state in GetState Add missing argument to NtQueryVirtualMemory_entry Fixed RtlCompareMemoryUlong_entry, it actually does not care if the source is misaligned, and for length it aligns down Fixed RtlUpperChar and RtlLowerChar, added a table that has their correct return values precomputed
2022-08-20 11:40:19 -07:00
parent 7cc364dcb8
commit 457296850e
54 changed files with 1441 additions and 615 deletions
--- a/src/xenia/cpu/backend/x64/x64_backend.cc
+++ b/src/xenia/cpu/backend/x64/x64_backend.cc
@@ -10,7 +10,7 @@
 #include "xenia/cpu/backend/x64/x64_backend.h"

 #include <stddef.h>
-
+#include <algorithm>
 #include "third_party/capstone/include/capstone/capstone.h"
 #include "third_party/capstone/include/capstone/x86.h"

@@ -50,6 +50,9 @@ DEFINE_bool(record_mmio_access_exceptions, true,
            "for them. This info can then be used on a subsequent run to "
            "instruct the recompiler to emit checks",
            "CPU");
+#if XE_X64_PROFILER_AVAILABLE == 1
+DECLARE_bool(instrument_call_times);
+#endif

 namespace xe {
 namespace cpu {
@@ -96,6 +99,68 @@ static void ForwardMMIOAccessForRecording(void* context, void* hostaddr) {
  reinterpret_cast<X64Backend*>(context)
      ->RecordMMIOExceptionForGuestInstruction(hostaddr);
 }
+#if XE_X64_PROFILER_AVAILABLE == 1
+// todo: better way of passing to atexit. maybe do in destructor instead?
+// nope, destructor is never called
+static GuestProfilerData* backend_profiler_data = nullptr;
+
+static uint64_t nanosecond_lifetime_start = 0;
+static void WriteGuestProfilerData() {
+  if (cvars::instrument_call_times) {
+    uint64_t end = Clock::QueryHostSystemTime();
+
+    uint64_t total = end - nanosecond_lifetime_start;
+
+    double totaltime_divisor = static_cast<double>(total);
+
+    FILE* output_file = nullptr;
+    std::vector<std::pair<uint32_t, uint64_t>> unsorted_profile{};
+    for (auto&& entry : *backend_profiler_data) {
+      if (entry.second) {  // skip times of 0
+        unsorted_profile.emplace_back(entry.first, entry.second);
+      }
+    }
+
+    std::sort(unsorted_profile.begin(), unsorted_profile.end(),
+              [](auto& x, auto& y) { return x.second < y.second; });
+
+    fopen_s(&output_file, "profile_times.txt", "w");
+    FILE* idapy_file = nullptr;
+    fopen_s(&idapy_file, "profile_print_times.py", "w");
+
+    for (auto&& sorted_entry : unsorted_profile) {
+      // double time_in_seconds =
+      //    static_cast<double>(sorted_entry.second) / 10000000.0;
+      double time_in_milliseconds =
+          static_cast<double>(sorted_entry.second) / (10000000.0 / 1000.0);
+
+      double slice = static_cast<double>(sorted_entry.second) /
+                     static_cast<double>(totaltime_divisor);
+
+      fprintf(output_file,
+              "%X took %.20f milliseconds, totaltime slice percentage %.20f \n",
+              sorted_entry.first, time_in_milliseconds, slice);
+
+      fprintf(idapy_file,
+              "print(get_name(0x%X) + ' took %.20f ms, %.20f percent')\n",
+              sorted_entry.first, time_in_milliseconds, slice);
+    }
+
+    fclose(output_file);
+    fclose(idapy_file);
+  }
+}
+
+static void GuestProfilerUpdateThreadProc() {
+  nanosecond_lifetime_start = Clock::QueryHostSystemTime();
+
+  do {
+    xe::threading::Sleep(std::chrono::seconds(30));
+    WriteGuestProfilerData();
+  } while (true);
+}
+static std::unique_ptr<xe::threading::Thread> g_profiler_update_thread{};
+#endif

 bool X64Backend::Initialize(Processor* processor) {
  if (!Backend::Initialize(processor)) {
@@ -159,6 +224,21 @@ bool X64Backend::Initialize(Processor* processor) {

  processor->memory()->SetMMIOExceptionRecordingCallback(
      ForwardMMIOAccessForRecording, (void*)this);
+
+#if XE_X64_PROFILER_AVAILABLE == 1
+  if (cvars::instrument_call_times) {
+    backend_profiler_data = &profiler_data_;
+    xe::threading::Thread::CreationParameters slimparams;
+
+    slimparams.create_suspended = false;
+    slimparams.initial_priority = xe::threading::ThreadPriority::kLowest;
+    slimparams.stack_size = 65536 * 4;
+
+    g_profiler_update_thread = std::move(xe::threading::Thread::Create(
+        slimparams, GuestProfilerUpdateThreadProc));
+  }
+#endif
+
  return true;
 }

@@ -734,6 +814,7 @@ void X64Backend::InitializeBackendContext(void* ctx) {
  bctx->flags = 0;
  // https://media.discordapp.net/attachments/440280035056943104/1000765256643125308/unknown.png
  bctx->Ox1000 = 0x1000;
+  bctx->guest_tick_count = Clock::GetGuestTickCountPointer();
 }
 const uint32_t mxcsr_table[8] = {
    0x1F80, 0x7F80, 0x5F80, 0x3F80, 0x9F80, 0xFF80, 0xDF80, 0xBF80,
@@ -747,6 +828,23 @@ void X64Backend::SetGuestRoundingMode(void* ctx, unsigned int mode) {
  bctx->mxcsr_fpu = mxcsr_table[control];
  ((ppc::PPCContext*)ctx)->fpscr.bits.rn = control;
 }
+
+#if XE_X64_PROFILER_AVAILABLE == 1
+uint64_t* X64Backend::GetProfilerRecordForFunction(uint32_t guest_address) {
+  // who knows, we might want to compile different versions of a function one
+  // day
+  auto entry = profiler_data_.find(guest_address);
+
+  if (entry != profiler_data_.end()) {
+    return &entry->second;
+  } else {
+    profiler_data_[guest_address] = 0;
+
+    return &profiler_data_[guest_address];
+  }
+}
+
+#endif
 }  // namespace x64
 }  // namespace backend
 }  // namespace cpu
--- a/src/xenia/cpu/backend/x64/x64_backend.h
+++ b/src/xenia/cpu/backend/x64/x64_backend.h
@@ -15,6 +15,14 @@
 #include "xenia/base/cvar.h"
 #include "xenia/cpu/backend/backend.h"

+#if XE_PLATFORM_WIN32 == 1
+// we use KUSER_SHARED's systemtime field, which is at a fixed address and
+// obviously windows specific, to get the start/end time for a function using
+// rdtsc would be too slow and skew the results by consuming extra cpu time, so
+// we have lower time precision but better overall accuracy
+#define XE_X64_PROFILER_AVAILABLE 1
+#endif
+
 DECLARE_int32(x64_extension_mask);

 namespace xe {
@@ -24,6 +32,8 @@ namespace xe {
 namespace cpu {
 namespace backend {
 namespace x64 {
+// mapping of guest function addresses to total nanoseconds taken in the func
+using GuestProfilerData = std::map<uint32_t, uint64_t>;

 class X64CodeCache;

@@ -37,8 +47,10 @@ typedef void (*ResolveFunctionThunk)();
 // negatively index the membase reg)
 struct X64BackendContext {
  void* ResolveFunction_Ptr;  // cached pointer to resolvefunction
-  unsigned int mxcsr_fpu;     // currently, the way we implement rounding mode
-                              // affects both vmx and the fpu
+  uint64_t* guest_tick_count;
+
+  unsigned int mxcsr_fpu;  // currently, the way we implement rounding mode
+                           // affects both vmx and the fpu
  unsigned int mxcsr_vmx;
  unsigned int flags;   // bit 0 = 0 if mxcsr is fpu, else it is vmx
  unsigned int Ox1000;  // constant 0x1000 so we can shrink each tail emitted
@@ -93,7 +105,9 @@ class X64Backend : public Backend {
  virtual void SetGuestRoundingMode(void* ctx, unsigned int mode) override;

  void RecordMMIOExceptionForGuestInstruction(void* host_address);
-
+#if XE_X64_PROFILER_AVAILABLE == 1
+  uint64_t* GetProfilerRecordForFunction(uint32_t guest_address);
+#endif
 private:
  static bool ExceptionCallbackThunk(Exception* ex, void* data);
  bool ExceptionCallback(Exception* ex);
@@ -106,6 +120,10 @@ class X64Backend : public Backend {
  HostToGuestThunk host_to_guest_thunk_;
  GuestToHostThunk guest_to_host_thunk_;
  ResolveFunctionThunk resolve_function_thunk_;
+
+#if XE_X64_PROFILER_AVAILABLE == 1
+  GuestProfilerData profiler_data_;
+#endif
 };

 }  // namespace x64
--- a/src/xenia/cpu/backend/x64/x64_emitter.cc
+++ b/src/xenia/cpu/backend/x64/x64_emitter.cc
@@ -57,6 +57,12 @@ DEFINE_bool(enable_incorrect_roundingmode_behavior, false,
            "code. The workaround may cause reduced CPU performance but is a "
            "more accurate emulation",
            "x64");
+
+#if XE_X64_PROFILER_AVAILABLE == 1
+DEFINE_bool(instrument_call_times, false,
+            "Compute time taken for functions, for profiling guest code",
+            "x64");
+#endif
 namespace xe {
 namespace cpu {
 namespace backend {
@@ -120,28 +126,37 @@ X64Emitter::X64Emitter(X64Backend* backend, XbyakAllocator* allocator)
 */
  unsigned int data[4];
  Xbyak::util::Cpu::getCpuid(0x80000001, data);
-  if (data[2] & (1U << 5)) {
+  unsigned amd_flags = data[2];
+  if (amd_flags & (1U << 5)) {
    if ((cvars::x64_extension_mask & kX64EmitLZCNT) == kX64EmitLZCNT) {
      feature_flags_ |= kX64EmitLZCNT;
    }
  }
+  // todo: although not reported by cpuid, zen 1 and zen+ also have fma4
+  if (amd_flags & (1U << 16)) {
+    if ((cvars::x64_extension_mask & kX64EmitFMA4) == kX64EmitFMA4) {
+      feature_flags_ |= kX64EmitFMA4;
+    }
+  }
+  if (amd_flags & (1U << 21)) {
+    if ((cvars::x64_extension_mask & kX64EmitTBM) == kX64EmitTBM) {
+      feature_flags_ |= kX64EmitTBM;
+    }
+  }
  if (cpu_.has(Xbyak::util::Cpu::tAMD)) {
    bool is_zennish = cpu_.displayFamily >= 0x17;
-
+    /*
+                chrispy: according to agner's tables, all amd architectures that
+       we support (ones with avx) have the same timings for
+       jrcxz/loop/loope/loopne as for other jmps
+        */
+    feature_flags_ |= kX64FastJrcx;
+    feature_flags_ |= kX64FastLoop;
    if (is_zennish) {
      // ik that i heard somewhere that this is the case for zen, but i need to
      // verify. cant find my original source for that.
      // todo: ask agner?
      feature_flags_ |= kX64FlagsIndependentVars;
-      feature_flags_ |= kX64FastJrcx;
-
-      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
    }
  }
  may_use_membase32_as_zero_reg_ =
@@ -157,6 +172,7 @@ bool X64Emitter::Emit(GuestFunction* function, HIRBuilder* builder,
                      std::vector<SourceMapEntry>* out_source_map) {
  SCOPE_profile_cpu_f("cpu");
  guest_module_ = dynamic_cast<XexModule*>(function->module());
+  current_guest_function_ = function->address();
  // Reset.
  debug_info_ = debug_info;
  debug_info_flags_ = debug_info_flags;
@@ -286,10 +302,19 @@ bool X64Emitter::Emit(HIRBuilder* builder, EmitFunctionInfo& func_info) {
  * 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], rax);  // 0

+#if XE_X64_PROFILER_AVAILABLE == 1
+  if (cvars::instrument_call_times) {
+    mov(rdx, 0x7ffe0014);  // load pointer to kusershared systemtime
+    mov(rdx, qword[rdx]);
+    mov(qword[rsp + StackLayout::GUEST_PROFILER_START],
+        rdx);  // save time for end of function
+  }
+#endif
  // Safe now to do some tracing.
  if (debug_info_flags_ & DebugInfoFlags::kDebugInfoTraceFunctions) {
    // We require 32-bit addresses.
@@ -363,6 +388,7 @@ bool X64Emitter::Emit(HIRBuilder* builder, EmitFunctionInfo& func_info) {
  mov(GetContextReg(), qword[rsp + StackLayout::GUEST_CTX_HOME]);
  */
  code_offsets.epilog = getSize();
+  EmitProfilerEpilogue();

  add(rsp, (uint32_t)stack_size);
  ret();
@@ -391,6 +417,27 @@ bool X64Emitter::Emit(HIRBuilder* builder, EmitFunctionInfo& func_info) {

  return true;
 }
+// dont use rax, we do this in tail call handling
+void X64Emitter::EmitProfilerEpilogue() {
+#if XE_X64_PROFILER_AVAILABLE == 1
+  if (cvars::instrument_call_times) {
+    uint64_t* profiler_entry =
+        backend()->GetProfilerRecordForFunction(current_guest_function_);
+    mov(ecx, 0x7ffe0014);
+    mov(rdx, qword[rcx]);
+    mov(rbx, (uintptr_t)profiler_entry);
+    sub(rdx, qword[rsp + StackLayout::GUEST_PROFILER_START]);
+
+    // atomic add our time to the profiler entry
+    // this could be atomic free if we had per thread profile counts, and on a
+    // threads exit we lock and sum up to the global counts, which would make
+    // this a few cycles less intrusive, but its good enough for now
+    //  actually... lets just try without atomics lol
+    // lock();
+    add(qword[rbx], rdx);
+  }
+#endif
+}

 void X64Emitter::MarkSourceOffset(const Instr* i) {
  auto entry = source_map_arena_.Alloc<SourceMapEntry>();
@@ -558,7 +605,7 @@ void X64Emitter::Call(const hir::Instr* instr, GuestFunction* function) {
  if (instr->flags & hir::CALL_TAIL) {
    // Since we skip the prolog we need to mark the return here.
    EmitTraceUserCallReturn();
-
+    EmitProfilerEpilogue();
    // Pass the callers return address over.
    mov(rcx, qword[rsp + StackLayout::GUEST_RET_ADDR]);

@@ -602,7 +649,7 @@ void X64Emitter::CallIndirect(const hir::Instr* instr,
  if (instr->flags & hir::CALL_TAIL) {
    // Since we skip the prolog we need to mark the return here.
    EmitTraceUserCallReturn();
-
+    EmitProfilerEpilogue();
    // Pass the callers return address over.
    mov(rcx, qword[rsp + StackLayout::GUEST_RET_ADDR]);

@@ -952,7 +999,34 @@ static const vec128_t xmm_consts[] = {
    /*XMMVSRShlByteshuf*/
    v128_setr_bytes(13, 14, 15, 8, 9, 10, 11, 4, 5, 6, 7, 0, 1, 2, 3, 0x80),
    // XMMVSRMask
-    vec128b(1)};
+    vec128b(1),
+    /*
+        XMMF16UnpackLCPI2
+        */
+
+    vec128i(0x38000000),
+    /*
+                XMMF16UnpackLCPI3
+        */
+    vec128q(0x7fe000007fe000ULL),
+
+    /* XMMF16PackLCPI0*/
+    vec128i(0x8000000),
+    /*XMMF16PackLCPI2*/
+    vec128i(0x47ffe000),
+    /*XMMF16PackLCPI3*/
+    vec128i(0xc7800000),
+    /*XMMF16PackLCPI4
+     */
+    vec128i(0xf7fdfff),
+    /*XMMF16PackLCPI5*/
+    vec128i(0x7fff),
+    /*
+    XMMF16PackLCPI6
+    */
+    vec128i(0x8000)
+
+};

 void* X64Emitter::FindByteConstantOffset(unsigned bytevalue) {
  for (auto& vec : xmm_consts) {
--- a/src/xenia/cpu/backend/x64/x64_emitter.h
+++ b/src/xenia/cpu/backend/x64/x64_emitter.h
@@ -159,7 +159,15 @@ enum XmmConst {
  XMMThreeFloatMask,  // for clearing the fourth float prior to DOT_PRODUCT_3
  XMMXenosF16ExtRangeStart,
  XMMVSRShlByteshuf,
-  XMMVSRMask
+  XMMVSRMask,
+  XMMF16UnpackLCPI2,  // 0x38000000, 1/ 32768
+  XMMF16UnpackLCPI3,  // 0x0x7fe000007fe000
+  XMMF16PackLCPI0,
+  XMMF16PackLCPI2,
+  XMMF16PackLCPI3,
+  XMMF16PackLCPI4,
+  XMMF16PackLCPI5,
+  XMMF16PackLCPI6
 };
 // X64Backend specific Instr->runtime_flags
 enum : uint32_t {
@@ -177,7 +185,7 @@ class XbyakAllocator : public Xbyak::Allocator {
 enum X64EmitterFeatureFlags {
  kX64EmitAVX2 = 1 << 0,
  kX64EmitFMA = 1 << 1,
-  kX64EmitLZCNT = 1 << 2,
+  kX64EmitLZCNT = 1 << 2,  // this is actually ABM and includes popcount
  kX64EmitBMI1 = 1 << 3,
  kX64EmitBMI2 = 1 << 4,
  kX64EmitF16C = 1 << 5,
@@ -201,7 +209,11 @@ enum X64EmitterFeatureFlags {
                // inc/dec) do not introduce false dependencies on EFLAGS
                // because the individual flags are treated as different vars by
                // the processor. (this applies to zen)
-  kX64EmitPrefetchW = 1 << 16
+  kX64EmitPrefetchW = 1 << 16,
+  kX64EmitXOP = 1 << 17,   // chrispy: xop maps really well to many vmx
+                           // instructions, and FX users need the boost
+  kX64EmitFMA4 = 1 << 18,  // todo: also use on zen1?
+  kX64EmitTBM = 1 << 19
 };
 class ResolvableGuestCall {
 public:
@@ -337,6 +349,8 @@ class X64Emitter : public Xbyak::CodeGenerator {

  XexModule* GuestModule() { return guest_module_; }

+  void EmitProfilerEpilogue();
+
 protected:
  void* Emplace(const EmitFunctionInfo& func_info,
                GuestFunction* function = nullptr);
@@ -352,7 +366,7 @@ class X64Emitter : public Xbyak::CodeGenerator {
  XexModule* guest_module_ = nullptr;
  Xbyak::util::Cpu cpu_;
  uint32_t feature_flags_ = 0;
-
+  uint32_t current_guest_function_ = 0;
  Xbyak::Label* epilog_label_ = nullptr;

  hir::Instr* current_instr_ = nullptr;
--- a/src/xenia/cpu/backend/x64/x64_seq_vector.cc
+++ b/src/xenia/cpu/backend/x64/x64_seq_vector.cc
@@ -19,10 +19,6 @@
 #include "xenia/base/cvar.h"
 #include "xenia/cpu/backend/x64/x64_stack_layout.h"

-DEFINE_bool(use_extended_range_half, true,
-            "Emulate extended range half-precision, may be slower on games "
-            "that use it heavily",
-            "CPU");
 namespace xe {
 namespace cpu {
 namespace backend {
@@ -1982,6 +1978,137 @@ 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) {
+  auto src1 = i.src1;
+
+  e.vpshufb(i.dest, src1, e.GetXmmConstPtr(initial_shuffle));
+  e.vpmovsxwd(e.xmm1, i.dest);
+
+  e.vpsrld(e.xmm2, e.xmm1, 10);
+  e.vpmovsxwd(e.xmm0, i.dest);
+  e.vpand(e.xmm0, e.xmm0, e.GetXmmConstPtr(XMMSignMaskPS));
+  e.vpand(e.xmm2, e.xmm2, e.GetXmmConstPtr(XMMPermuteByteMask));
+
+  e.vpslld(e.xmm3, e.xmm2, 23);
+
+  e.vpaddd(e.xmm3, e.xmm3, e.GetXmmConstPtr(XMMF16UnpackLCPI2));
+
+  e.vpcmpeqd(e.xmm2, e.xmm2, e.GetXmmConstPtr(XMMZero));
+
+  e.vpslld(e.xmm1, e.xmm1, 13);
+
+  e.vpandn(e.xmm1, e.xmm2, e.xmm1);
+  e.vpandn(e.xmm2, e.xmm2, e.xmm3);
+
+  e.vpand(e.xmm1, e.xmm1, e.GetXmmConstPtr(XMMF16UnpackLCPI3));
+  e.vpor(e.xmm0, e.xmm1, e.xmm0);
+  e.vpor(i.dest, e.xmm0, e.xmm2);
+}
+template <typename Inst>
+static void emit_fast_f16_pack(X64Emitter& e, const Inst& i,
+                               XmmConst final_shuffle) {
+  e.vpaddd(e.xmm1, i.src1, e.GetXmmConstPtr(XMMF16PackLCPI0));
+  e.vpand(e.xmm2, i.src1, e.GetXmmConstPtr(XMMAbsMaskPS));
+  e.vmovdqa(e.xmm3, e.GetXmmConstPtr(XMMF16PackLCPI2));
+
+  e.vpcmpgtd(e.xmm3, e.xmm3, e.xmm2);
+  e.vpsrld(e.xmm1, e.xmm1, 13);
+
+  e.vpaddd(e.xmm2, e.xmm2, e.GetXmmConstPtr(XMMF16PackLCPI3));
+  e.vpminud(e.xmm0, e.xmm2, e.GetXmmConstPtr(XMMF16PackLCPI4));
+
+  e.vpcmpeqd(e.xmm2, e.xmm2, e.xmm0);
+  e.vmovdqa(e.xmm0, e.GetXmmConstPtr(XMMF16PackLCPI5));
+  e.vpand(e.xmm1, e.xmm1, e.xmm0);
+  e.vpand(e.xmm1, e.xmm2, e.xmm1);
+  e.vpxor(e.xmm2, e.xmm2, e.xmm2);
+
+  e.vblendvps(e.xmm1, e.xmm0, e.xmm1, e.xmm3);
+
+  e.vpsrld(e.xmm0, i.src1, 16);
+  e.vpand(e.xmm0, e.xmm0, e.GetXmmConstPtr(XMMF16PackLCPI6));
+  e.vorps(e.xmm0, e.xmm1, e.xmm0);
+  e.vpackusdw(i.dest, e.xmm0, e.xmm2);
+  e.vpshufb(i.dest, i.dest, e.GetXmmConstPtr(final_shuffle));
+}
 // ============================================================================
 // OPCODE_SWIZZLE
 // ============================================================================
@@ -2081,14 +2208,9 @@ struct PACK : Sequence<PACK, I<OPCODE_PACK, V128Op, V128Op, V128Op>> {
    alignas(16) uint16_t b[8];
    _mm_store_ps(a, src1);
    std::memset(b, 0, sizeof(b));
-    if (!cvars::use_extended_range_half) {
-      for (int i = 0; i < 2; i++) {
-        b[7 - i] = half_float::detail::float2half<std::round_toward_zero>(a[i]);
-      }
-    } else {
-      for (int i = 0; i < 2; i++) {
-        b[7 - i] = float_to_xenos_half(a[i]);
-      }
+
+    for (int i = 0; i < 2; i++) {
+      b[7 - i] = float_to_xenos_half(a[i]);
    }

    return _mm_load_si128(reinterpret_cast<__m128i*>(b));
@@ -2098,70 +2220,26 @@ struct PACK : Sequence<PACK, I<OPCODE_PACK, V128Op, V128Op, V128Op>> {
    // http://blogs.msdn.com/b/chuckw/archive/2012/09/11/directxmath-f16c-and-fma.aspx
    // dest = [(src1.x | src1.y), 0, 0, 0]

-    if (e.IsFeatureEnabled(kX64EmitF16C) && !cvars::use_extended_range_half) {
-      Xmm src;
-      if (i.src1.is_constant) {
-        src = i.dest;
-        e.LoadConstantXmm(src, i.src1.constant());
-      } else {
-        src = i.src1;
-      }
-      // 0|0|0|0|W|Z|Y|X
-      e.vcvtps2ph(i.dest, src, 0b00000011);
-      // Shuffle to X|Y|0|0|0|0|0|0
-      e.vpshufb(i.dest, i.dest, e.GetXmmConstPtr(XMMPackFLOAT16_2));
+    if (i.src1.is_constant) {
+      e.lea(e.GetNativeParam(0), e.StashConstantXmm(0, i.src1.constant()));
    } else {
-      if (i.src1.is_constant) {
-        e.lea(e.GetNativeParam(0), e.StashConstantXmm(0, i.src1.constant()));
-      } else {
-        e.lea(e.GetNativeParam(0), e.StashXmm(0, i.src1));
-      }
-      e.CallNativeSafe(reinterpret_cast<void*>(EmulateFLOAT16_2));
-      e.vmovaps(i.dest, e.xmm0);
+      e.lea(e.GetNativeParam(0), e.StashXmm(0, i.src1));
    }
+    e.CallNativeSafe(reinterpret_cast<void*>(EmulateFLOAT16_2));
+    e.vmovaps(i.dest, e.xmm0);
  }
-  static __m128i EmulateFLOAT16_4(void*, __m128 src1) {
-    alignas(16) float a[4];
-    alignas(16) uint16_t b[8];
-    _mm_store_ps(a, src1);
-    std::memset(b, 0, sizeof(b));
-    if (!cvars::use_extended_range_half) {
-      for (int i = 0; i < 4; i++) {
-        b[7 - (i ^ 2)] =
-            half_float::detail::float2half<std::round_toward_zero>(a[i]);
-      }
-    } else {
-      for (int i = 0; i < 4; i++) {
-        b[7 - (i ^ 2)] = float_to_xenos_half(a[i]);
-      }
-    }

-    return _mm_load_si128(reinterpret_cast<__m128i*>(b));
-  }
  static void EmitFLOAT16_4(X64Emitter& e, const EmitArgType& i) {
-    assert_true(i.src2.value->IsConstantZero());
-    // dest = [(src1.z | src1.w), (src1.x | src1.y), 0, 0]
-
-    if (e.IsFeatureEnabled(kX64EmitF16C) && !cvars::use_extended_range_half) {
-      Xmm src;
-      if (i.src1.is_constant) {
-        src = i.dest;
-        e.LoadConstantXmm(src, i.src1.constant());
-      } else {
-        src = i.src1;
-      }
-      // 0|0|0|0|W|Z|Y|X
-      e.vcvtps2ph(i.dest, src, 0b00000011);
-      // Shuffle to Z|W|X|Y|0|0|0|0
-      e.vpshufb(i.dest, i.dest, e.GetXmmConstPtr(XMMPackFLOAT16_4));
+    if (!i.src1.is_constant) {
+      emit_fast_f16_pack(e, i, XMMPackFLOAT16_4);
    } else {
-      if (i.src1.is_constant) {
-        e.lea(e.GetNativeParam(0), e.StashConstantXmm(0, i.src1.constant()));
-      } else {
-        e.lea(e.GetNativeParam(0), e.StashXmm(0, i.src1));
+      vec128_t result = vec128b(0);
+      for (unsigned idx = 0; idx < 4; ++idx) {
+        result.u16[(7 - (idx ^ 2))] =
+            float_to_xenos_half(i.src1.constant().f32[idx]);
      }
-      e.CallNativeSafe(reinterpret_cast<void*>(EmulateFLOAT16_4));
-      e.vmovaps(i.dest, e.xmm0);
+
+      e.LoadConstantXmm(i.dest, result);
    }
  }
  static void EmitSHORT_2(X64Emitter& e, const EmitArgType& i) {
@@ -2508,15 +2586,10 @@ struct UNPACK : Sequence<UNPACK, I<OPCODE_UNPACK, V128Op, V128Op>> {
    alignas(16) float b[4];
    _mm_store_si128(reinterpret_cast<__m128i*>(a), src1);

-    if (!cvars::use_extended_range_half) {
-      for (int i = 0; i < 2; i++) {
-        b[i] = half_float::detail::half2float(a[VEC128_W(6 + i)]);
-      }
-    } else {
-      for (int i = 0; i < 2; i++) {
-        b[i] = xenos_half_to_float(a[VEC128_W(6 + i)]);
-      }
+    for (int i = 0; i < 2; i++) {
+      b[i] = xenos_half_to_float(a[VEC128_W(6 + i)]);
    }
+
    // Constants, or something
    b[2] = 0.f;
    b[3] = 1.f;
@@ -2536,74 +2609,28 @@ struct UNPACK : Sequence<UNPACK, I<OPCODE_UNPACK, V128Op, V128Op>> {
    // Also zero out the high end.
    // TODO(benvanik): special case constant unpacks that just get 0/1/etc.

-    if (e.IsFeatureEnabled(kX64EmitF16C) &&
-        !cvars::use_extended_range_half) {  // todo: can use cvtph and bit logic
-                                            // to implement
-      Xmm src;
-      if (i.src1.is_constant) {
-        src = i.dest;
-        e.LoadConstantXmm(src, i.src1.constant());
-      } else {
-        src = i.src1;
-      }
-      // sx = src.iw >> 16;
-      // sy = src.iw & 0xFFFF;
-      // dest = { XMConvertHalfToFloat(sx),
-      //          XMConvertHalfToFloat(sy),
-      //          0.0,
-      //          1.0 };
-      // Shuffle to 0|0|0|0|0|0|Y|X
-      e.vpshufb(i.dest, src, e.GetXmmConstPtr(XMMUnpackFLOAT16_2));
-      e.vcvtph2ps(i.dest, i.dest);
-      e.vpshufd(i.dest, i.dest, 0b10100100);
-      e.vpor(i.dest, e.GetXmmConstPtr(XMM0001));
+    if (i.src1.is_constant) {
+      e.lea(e.GetNativeParam(0), e.StashConstantXmm(0, i.src1.constant()));
    } else {
-      if (i.src1.is_constant) {
-        e.lea(e.GetNativeParam(0), e.StashConstantXmm(0, i.src1.constant()));
-      } else {
-        e.lea(e.GetNativeParam(0), e.StashXmm(0, i.src1));
-      }
-      e.CallNativeSafe(reinterpret_cast<void*>(EmulateFLOAT16_2));
-      e.vmovaps(i.dest, e.xmm0);
+      e.lea(e.GetNativeParam(0), e.StashXmm(0, i.src1));
    }
+    e.CallNativeSafe(reinterpret_cast<void*>(EmulateFLOAT16_2));
+    e.vmovaps(i.dest, e.xmm0);
  }
-  static __m128 EmulateFLOAT16_4(void*, __m128i src1) {
-    alignas(16) uint16_t a[8];
-    alignas(16) float b[4];
-    _mm_store_si128(reinterpret_cast<__m128i*>(a), src1);
-    if (!cvars::use_extended_range_half) {
-      for (int i = 0; i < 4; i++) {
-        b[i] = half_float::detail::half2float(a[VEC128_W(4 + i)]);
-      }
-    } else {
-      for (int i = 0; i < 4; i++) {
-        b[i] = xenos_half_to_float(a[VEC128_W(4 + i)]);
-      }
-    }

-    return _mm_load_ps(b);
-  }
  static void EmitFLOAT16_4(X64Emitter& e, const EmitArgType& i) {
-    // src = [(dest.x | dest.y), (dest.z | dest.w), 0, 0]
-    if (e.IsFeatureEnabled(kX64EmitF16C) && !cvars::use_extended_range_half) {
-      Xmm src;
-      if (i.src1.is_constant) {
-        src = i.dest;
-        e.LoadConstantXmm(src, i.src1.constant());
-      } else {
-        src = i.src1;
+    if (i.src1.is_constant) {
+      vec128_t result{};
+
+      for (int idx = 0; idx < 4; ++idx) {
+        result.f32[idx] =
+            xenos_half_to_float(i.src1.constant().u16[VEC128_W(4 + idx)]);
      }
-      // Shuffle to 0|0|0|0|W|Z|Y|X
-      e.vpshufb(i.dest, src, e.GetXmmConstPtr(XMMUnpackFLOAT16_4));
-      e.vcvtph2ps(i.dest, i.dest);
+
+      e.LoadConstantXmm(i.dest, result);
+
    } else {
-      if (i.src1.is_constant) {
-        e.lea(e.GetNativeParam(0), e.StashConstantXmm(0, i.src1.constant()));
-      } else {
-        e.lea(e.GetNativeParam(0), e.StashXmm(0, i.src1));
-      }
-      e.CallNativeSafe(reinterpret_cast<void*>(EmulateFLOAT16_4));
-      e.vmovaps(i.dest, e.xmm0);
+      emit_fast_f16_unpack(e, i, XMMUnpackFLOAT16_4);
    }
  }
  static void EmitSHORT_2(X64Emitter& e, const EmitArgType& i) {
--- a/src/xenia/cpu/backend/x64/x64_sequences.cc
+++ b/src/xenia/cpu/backend/x64/x64_sequences.cc
@@ -50,6 +50,10 @@ DEFINE_bool(no_round_to_single, false,
            "Not for users, breaks games. Skip rounding double values to "
            "single precision and back",
            "CPU");
+DEFINE_bool(
+    inline_loadclock, false,
+    "Directly read cached guest clock without calling the LoadClock method (it gets repeatedly updated by calls from other threads)",
+    "CPU");
 namespace xe {
 namespace cpu {
 namespace backend {
@@ -475,33 +479,39 @@ EMITTER_OPCODE_TABLE(OPCODE_ROUND, ROUND_F32, ROUND_F64, ROUND_V128);
 // ============================================================================
 struct LOAD_CLOCK : Sequence<LOAD_CLOCK, I<OPCODE_LOAD_CLOCK, I64Op>> {
  static void Emit(X64Emitter& e, const EmitArgType& i) {
-    // When scaling is disabled and the raw clock source is selected, the code
-    // in the Clock class is actually just forwarding tick counts after one
-    // simple multiply and division. In that case we rather bake the scaling in
-    // here to cut extra function calls with CPU cache misses and stack frame
-    // overhead.
-    if (cvars::clock_no_scaling && cvars::clock_source_raw) {
-      auto ratio = Clock::guest_tick_ratio();
-      // The 360 CPU is an in-order CPU, AMD64 usually isn't. Without
-      // mfence/lfence magic the rdtsc instruction can be executed sooner or
-      // later in the cache window. Since it's resolution however is much higher
-      // than the 360's mftb instruction this can safely be ignored.
-
-      // Read time stamp in edx (high part) and eax (low part).
-      e.rdtsc();
-      // Make it a 64 bit number in rax.
-      e.shl(e.rdx, 32);
-      e.or_(e.rax, e.rdx);
-      // Apply tick frequency scaling.
-      e.mov(e.rcx, ratio.first);
-      e.mul(e.rcx);
-      // We actually now have a 128 bit number in rdx:rax.
-      e.mov(e.rcx, ratio.second);
-      e.div(e.rcx);
-      e.mov(i.dest, e.rax);
+    if (cvars::inline_loadclock) {
+      e.mov(e.rcx,
+            e.GetBackendCtxPtr(offsetof(X64BackendContext, guest_tick_count)));
+      e.mov(i.dest, e.qword[e.rcx]);
    } else {
-      e.CallNative(LoadClock);
-      e.mov(i.dest, e.rax);
+      // When scaling is disabled and the raw clock source is selected, the code
+      // in the Clock class is actually just forwarding tick counts after one
+      // simple multiply and division. In that case we rather bake the scaling
+      // in here to cut extra function calls with CPU cache misses and stack
+      // frame overhead.
+      if (cvars::clock_no_scaling && cvars::clock_source_raw) {
+        auto ratio = Clock::guest_tick_ratio();
+        // The 360 CPU is an in-order CPU, AMD64 usually isn't. Without
+        // mfence/lfence magic the rdtsc instruction can be executed sooner or
+        // later in the cache window. Since it's resolution however is much
+        // higher than the 360's mftb instruction this can safely be ignored.
+
+        // Read time stamp in edx (high part) and eax (low part).
+        e.rdtsc();
+        // Make it a 64 bit number in rax.
+        e.shl(e.rdx, 32);
+        e.or_(e.rax, e.rdx);
+        // Apply tick frequency scaling.
+        e.mov(e.rcx, ratio.first);
+        e.mul(e.rcx);
+        // We actually now have a 128 bit number in rdx:rax.
+        e.mov(e.rcx, ratio.second);
+        e.div(e.rcx);
+        e.mov(i.dest, e.rax);
+      } else {
+        e.CallNative(LoadClock);
+        e.mov(i.dest, e.rax);
+      }
    }
  }
  static uint64_t LoadClock(void* raw_context) {
@@ -539,10 +549,12 @@ struct MAX_F64 : Sequence<MAX_F64, I<OPCODE_MAX, F64Op, F64Op, F64Op>> {
 struct MAX_V128 : Sequence<MAX_V128, I<OPCODE_MAX, V128Op, V128Op, V128Op>> {
  static void Emit(X64Emitter& e, const EmitArgType& i) {
    e.ChangeMxcsrMode(MXCSRMode::Vmx);
-    EmitCommutativeBinaryXmmOp(e, i,
-                               [](X64Emitter& e, Xmm dest, Xmm src1, Xmm src2) {
-                                 e.vmaxps(dest, src1, src2);
-                               });
+
+    auto src1 = GetInputRegOrConstant(e, i.src1, e.xmm0);
+    auto src2 = GetInputRegOrConstant(e, i.src2, e.xmm1);
+    e.vmaxps(e.xmm2, src1, src2);
+    e.vmaxps(e.xmm3, src2, src1);
+    e.vorps(i.dest, e.xmm2, e.xmm3);
  }
 };
 EMITTER_OPCODE_TABLE(OPCODE_MAX, MAX_F32, MAX_F64, MAX_V128);
@@ -597,10 +609,11 @@ struct MIN_F64 : Sequence<MIN_F64, I<OPCODE_MIN, F64Op, F64Op, F64Op>> {
 struct MIN_V128 : Sequence<MIN_V128, I<OPCODE_MIN, V128Op, V128Op, V128Op>> {
  static void Emit(X64Emitter& e, const EmitArgType& i) {
    e.ChangeMxcsrMode(MXCSRMode::Vmx);
-    EmitCommutativeBinaryXmmOp(e, i,
-                               [](X64Emitter& e, Xmm dest, Xmm src1, Xmm src2) {
-                                 e.vminps(dest, src1, src2);
-                               });
+    auto src1 = GetInputRegOrConstant(e, i.src1, e.xmm0);
+    auto src2 = GetInputRegOrConstant(e, i.src2, e.xmm1);
+    e.vminps(e.xmm2, src1, src2);
+    e.vminps(e.xmm3, src2, src1);
+    e.vorps(i.dest, e.xmm2, e.xmm3);
  }
 };
 EMITTER_OPCODE_TABLE(OPCODE_MIN, MIN_I8, MIN_I16, MIN_I32, MIN_I64, MIN_F32,
@@ -768,6 +781,7 @@ struct SELECT_V128_V128
    } else if (mayblend == PermittedBlend::Ps) {
      e.vblendvps(i.dest, src2, src3, src1);
    } else {
+		//ideally we would have an xop path here...
      // src1 ? src2 : src3;
      e.vpandn(e.xmm3, src1, src2);
      e.vpand(i.dest, src1, src3);
@@ -1932,6 +1946,53 @@ struct MUL_ADD_V128
 };
 EMITTER_OPCODE_TABLE(OPCODE_MUL_ADD, MUL_ADD_F32, MUL_ADD_F64, MUL_ADD_V128);

+struct NEGATED_MUL_ADD_F64
+    : Sequence<NEGATED_MUL_ADD_F64,
+               I<OPCODE_NEGATED_MUL_ADD, F64Op, F64Op, F64Op, F64Op>> {
+  static void Emit(X64Emitter& e, const EmitArgType& i) {
+    e.ChangeMxcsrMode(MXCSRMode::Fpu);
+
+    Xmm src1 = GetInputRegOrConstant(e, i.src1, e.xmm0);
+    Xmm src2 = GetInputRegOrConstant(e, i.src2, e.xmm1);
+    Xmm src3 = GetInputRegOrConstant(e, i.src3, e.xmm2);
+    if (e.IsFeatureEnabled(kX64EmitFMA)) {
+      // todo: this is garbage
+      e.vmovapd(e.xmm3, src1);
+      e.vfmadd213sd(e.xmm3, src2, src3);
+      e.vxorpd(i.dest, e.xmm3, e.GetXmmConstPtr(XMMSignMaskPD));
+    } else {
+      // todo: might need to use x87 in this case...
+      e.vmulsd(e.xmm3, src1, src2);
+      e.vaddsd(i.dest, e.xmm3, src3);
+      e.vxorpd(i.dest, i.dest, e.GetXmmConstPtr(XMMSignMaskPD));
+    }
+  }
+};
+struct NEGATED_MUL_ADD_V128
+    : Sequence<NEGATED_MUL_ADD_V128,
+               I<OPCODE_NEGATED_MUL_ADD, V128Op, V128Op, V128Op, V128Op>> {
+  static void Emit(X64Emitter& e, const EmitArgType& i) {
+    e.ChangeMxcsrMode(MXCSRMode::Vmx);
+
+    Xmm src1 = GetInputRegOrConstant(e, i.src1, e.xmm0);
+    Xmm src2 = GetInputRegOrConstant(e, i.src2, e.xmm1);
+    Xmm src3 = GetInputRegOrConstant(e, i.src3, e.xmm2);
+    if (e.IsFeatureEnabled(kX64EmitFMA)) {
+      // todo: this is garbage
+      e.vmovaps(e.xmm3, src1);
+      e.vfmadd213ps(e.xmm3, src2, src3);
+      e.vxorps(i.dest, e.xmm3, e.GetXmmConstPtr(XMMSignMaskPS));
+    } else {
+      // todo: might need to use x87 in this case...
+      e.vmulps(e.xmm3, src1, src2);
+      e.vaddps(i.dest, e.xmm3, src3);
+      e.vxorps(i.dest, i.dest, e.GetXmmConstPtr(XMMSignMaskPS));
+    }
+  }
+};
+EMITTER_OPCODE_TABLE(OPCODE_NEGATED_MUL_ADD, NEGATED_MUL_ADD_F64,
+                     NEGATED_MUL_ADD_V128);
+
 // ============================================================================
 // OPCODE_MUL_SUB
 // ============================================================================
@@ -1944,12 +2005,7 @@ EMITTER_OPCODE_TABLE(OPCODE_MUL_ADD, MUL_ADD_F32, MUL_ADD_F64, MUL_ADD_V128);
 // - 132 -> $1 = $1 * $3 - $2
 // - 213 -> $1 = $2 * $1 - $3
 // - 231 -> $1 = $2 * $3 - $1
-struct MUL_SUB_F32
-    : Sequence<MUL_SUB_F32, I<OPCODE_MUL_SUB, F32Op, F32Op, F32Op, F32Op>> {
-  static void Emit(X64Emitter& e, const EmitArgType& i) {
-    assert_impossible_sequence(MUL_SUB_F32);
-  }
-};
+
 struct MUL_SUB_F64
    : Sequence<MUL_SUB_F64, I<OPCODE_MUL_SUB, F64Op, F64Op, F64Op, F64Op>> {
  static void Emit(X64Emitter& e, const EmitArgType& i) {
@@ -1991,7 +2047,54 @@ struct MUL_SUB_V128
    }
  }
 };
-EMITTER_OPCODE_TABLE(OPCODE_MUL_SUB, MUL_SUB_F32, MUL_SUB_F64, MUL_SUB_V128);
+EMITTER_OPCODE_TABLE(OPCODE_MUL_SUB, MUL_SUB_F64, MUL_SUB_V128);
+
+struct NEGATED_MUL_SUB_F64
+    : Sequence<NEGATED_MUL_SUB_F64,
+               I<OPCODE_NEGATED_MUL_SUB, F64Op, F64Op, F64Op, F64Op>> {
+  static void Emit(X64Emitter& e, const EmitArgType& i) {
+    e.ChangeMxcsrMode(MXCSRMode::Fpu);
+
+    Xmm src1 = GetInputRegOrConstant(e, i.src1, e.xmm0);
+    Xmm src2 = GetInputRegOrConstant(e, i.src2, e.xmm1);
+    Xmm src3 = GetInputRegOrConstant(e, i.src3, e.xmm2);
+    if (e.IsFeatureEnabled(kX64EmitFMA)) {
+      // todo: this is garbage
+      e.vmovapd(e.xmm3, src1);
+      e.vfmsub213sd(e.xmm3, src2, src3);
+      e.vxorpd(i.dest, e.xmm3, e.GetXmmConstPtr(XMMSignMaskPD));
+    } else {
+      // todo: might need to use x87 in this case...
+      e.vmulsd(e.xmm3, src1, src2);
+      e.vsubsd(i.dest, e.xmm3, src3);
+      e.vxorpd(i.dest, i.dest, e.GetXmmConstPtr(XMMSignMaskPD));
+    }
+  }
+};
+struct NEGATED_MUL_SUB_V128
+    : Sequence<NEGATED_MUL_SUB_V128,
+               I<OPCODE_NEGATED_MUL_SUB, V128Op, V128Op, V128Op, V128Op>> {
+  static void Emit(X64Emitter& e, const EmitArgType& i) {
+    e.ChangeMxcsrMode(MXCSRMode::Vmx);
+
+    Xmm src1 = GetInputRegOrConstant(e, i.src1, e.xmm0);
+    Xmm src2 = GetInputRegOrConstant(e, i.src2, e.xmm1);
+    Xmm src3 = GetInputRegOrConstant(e, i.src3, e.xmm2);
+    if (e.IsFeatureEnabled(kX64EmitFMA)) {
+      // todo: this is garbage
+      e.vmovaps(e.xmm3, src1);
+      e.vfmsub213ps(e.xmm3, src2, src3);
+      e.vxorps(i.dest, e.xmm3, e.GetXmmConstPtr(XMMSignMaskPS));
+    } else {
+      // todo: might need to use x87 in this case...
+      e.vmulps(e.xmm3, src1, src2);
+      e.vsubps(i.dest, e.xmm3, src3);
+      e.vxorps(i.dest, i.dest, e.GetXmmConstPtr(XMMSignMaskPS));
+    }
+  }
+};
+EMITTER_OPCODE_TABLE(OPCODE_NEGATED_MUL_SUB, NEGATED_MUL_SUB_F64,
+                     NEGATED_MUL_SUB_V128);

 // ============================================================================
 // OPCODE_NEG
@@ -2264,7 +2367,7 @@ struct DOT_PRODUCT_3_V128
    e.ChangeMxcsrMode(MXCSRMode::Vmx);
    // 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];
+    auto mxcsr_storage = e.dword[e.rsp + StackLayout::GUEST_SCRATCH];

    // this is going to hurt a bit...
    /*
@@ -2380,7 +2483,7 @@ struct DOT_PRODUCT_4_V128
    e.ChangeMxcsrMode(MXCSRMode::Vmx);
    // 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];
+    auto mxcsr_storage = e.dword[e.rsp + StackLayout::GUEST_SCRATCH];

    bool is_lensqr = i.instr->src1.value == i.instr->src2.value;

@@ -3162,9 +3265,9 @@ struct SET_ROUNDING_MODE_I32
    // backends dont have to worry about it
    if (i.src1.is_constant) {
      e.mov(e.eax, mxcsr_table[i.src1.constant()]);
-      e.mov(e.dword[e.rsp + StackLayout::GUEST_SCRATCH64], e.eax);
+      e.mov(e.dword[e.rsp + StackLayout::GUEST_SCRATCH], e.eax);
      e.mov(e.GetBackendCtxPtr(offsetof(X64BackendContext, mxcsr_fpu)), e.eax);
-      e.vldmxcsr(e.dword[e.rsp + StackLayout::GUEST_SCRATCH64]);
+      e.vldmxcsr(e.dword[e.rsp + StackLayout::GUEST_SCRATCH]);

    } else {
      e.mov(e.ecx, i.src1);
--- a/src/xenia/cpu/backend/x64/x64_stack_layout.h
+++ b/src/xenia/cpu/backend/x64/x64_stack_layout.h
@@ -123,7 +123,10 @@ class StackLayout {
   */
  static const size_t GUEST_STACK_SIZE = 104;
  //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_SCRATCH = 0;
+  
+  //when profiling is on, this stores the nanosecond time at the start of the function
+  static const size_t GUEST_PROFILER_START = 80;
  static const size_t GUEST_RET_ADDR = 88;
  static const size_t GUEST_CALL_RET_ADDR = 96;
 };