Huge set of performance improvements, combined with an architecture specific build and clang-cl users have reported absurd gains over master for some gains, in the range 50%-90%

But for normal msvc builds i would put it at around 30-50%
Added per-xexmodule caching of information per instruction, can be used to remember what code needs compiling at start up
Record what guest addresses wrote mmio and backpropagate that to future runs, eliminating dependence on exception trapping. this makes many games like h3 actually tolerable to run under a debugger
fixed a number of errors where temporaries were being passed by reference/pointer
Can now be compiled with clang-cl 14.0.1, requires -Werror off though and some other solution/project changes.
Added macros wrapping compiler extensions like noinline, forceinline, __expect, and cold.
Removed the "global lock" in guest code completely. It does not properly emulate the behavior of mfmsrd/mtmsr and it seriously cripples amd cpus. Removing this yielded around a 3x speedup in Halo Reach for me.
Disabled the microprofiler for now. The microprofiler has a huge performance cost associated with it. Developers can re-enable it in the base/profiling header if they really need it
Disable the trace writer in release builds. despite just returning after checking if the file was open the trace functions were consuming about 0.60% cpu time total
Add IsValidReg, GetRegisterInfo is a huge (about 45k) branching function and using that to check if a register was valid consumed a significant chunk of time
Optimized RingBuffer::ReadAndSwap and RingBuffer::read_count. This gave us the largest overall boost in performance. The memcpies were unnecessary and one of them was always a no-op
Added simplification rules for multiplicative patterns like (x+x), (x<<1)+x
For the most frequently called win32 functions i added code to call their underlying NT implementations, which lets us skip a lot of MS code we don't care about/isnt relevant to our usecases
^this can be toggled off in the platform_win header
handle indirect call true with constant function pointer, was occurring in h3
lookup host format swizzle in denser array
by default, don't check if a gpu register is unknown, instead just check if its out of range. controlled by a cvar
^looking up whether its known or not took approx 0.3% cpu time
Changed some things in /cpu to make the project UNITYBUILD friendly
The timer thread was spinning way too much and consuming a ton of cpu, changed it to use a blocking wait instead
tagged some conditions as XE_UNLIKELY/LIKELY based on profiler feedback (will only affect clang builds)
Shifted around some code in CommandProcessor::WriteRegister based on how frequently it was executed
added support for docdecaduple precision floating point so that we can represent our performance gains numerically
tons of other stuff im probably forgetting

src/xenia/base/ring_buffer.h

@@ -19,7 +19,26 @@
 #include "xenia/base/byte_order.h"
 
 namespace xe {
+/*
+        todo: this class is CRITICAL to the performance of the entire emulator
+        currently, about 0.74% cpu time is still taken up by ReadAndSwap, 0.23
+   is used by read_count I believe that part of the issue is that smaller
+   ringbuffers are kicking off an automatic prefetcher stream, that ends up
+   reading ahead of the end of the ring because it can only go in a straight
+   line it then gets a cache miss when it eventually wraps around to the start
+   of the ring? really hard to tell whats going on there honestly, maybe we can
+   occasionally prefetch the first line of the ring to L1? For the automatic
+   prefetching i don't think there are any good options. I don't know if we have
+   any control over where these buffers will be (they seem to be in guest memory
+   :/), but if we did we could right-justify the buffer so that the final byte
+   of the ring ends at the end of a page. i think most automatic prefetchers
+   cannot cross page boundaries it does feel like something isnt right here
+   though
 
+        todo: microoptimization, we can change our size members to be uint32 so
+   that the registers no longer need the rex prefix, shrinking the generated
+   code a bit.. like i said, every bit helps in this class
+*/
 class RingBuffer {
  public:
   RingBuffer(uint8_t* buffer, size_t capacity);
@@ -32,6 +51,8 @@ class RingBuffer {
   uintptr_t read_ptr() const { return uintptr_t(buffer_) + read_offset_; }
   void set_read_offset(size_t offset) { read_offset_ = offset % capacity_; }
   size_t read_count() const {
+// chrispy: these branches are unpredictable
+#if 0
     if (read_offset_ == write_offset_) {
       return 0;
     } else if (read_offset_ < write_offset_) {
@@ -39,6 +60,33 @@ class RingBuffer {
     } else {
       return (capacity_ - read_offset_) + write_offset_;
     }
+#else
+    size_t read_offs = read_offset_;
+    size_t write_offs = write_offset_;
+    size_t cap = capacity_;
+
+    size_t offset_delta = write_offs - read_offs;
+    size_t wrap_read_count = (cap - read_offs) + write_offs;
+
+    size_t comparison_value = read_offs <= write_offs;
+#if 0
+    size_t selector =
+        static_cast<size_t>(-static_cast<ptrdiff_t>(comparison_value));
+    offset_delta &= selector;
+
+    wrap_read_count &= ~selector;
+    return offset_delta | wrap_read_count;
+#else
+
+    if (XE_LIKELY(read_offs <= write_offs)) {
+      return offset_delta;  // will be 0 if they are equal, semantically
+                            // identical to old code (i checked the asm, msvc
+                            // does not automatically do this)
+    } else {
+      return wrap_read_count;
+    }
+#endif
+#endif
   }
 
   size_t write_offset() const { return write_offset_; }
@@ -113,6 +161,28 @@ class RingBuffer {
   size_t write_offset_ = 0;
 };
 
+template <>
+inline uint32_t RingBuffer::ReadAndSwap<uint32_t>() {
+  size_t read_offset = this->read_offset_;
+  xenia_assert(this->capacity_ >= 4);
+
+  size_t next_read_offset = read_offset + 4;
+  #if 0
+  size_t zerotest = next_read_offset - this->capacity_;
+  // unpredictable branch, use bit arith instead
+  // todo: it would be faster to use lzcnt, but we need to figure out if all
+  // machines we support support it
+  next_read_offset &= -static_cast<ptrdiff_t>(!!zerotest);
+  #else
+  if (XE_UNLIKELY(next_read_offset == this->capacity_)) {
+    next_read_offset = 0;
+	//todo: maybe prefetch next? or should that happen much earlier?
+  }
+  #endif
+  this->read_offset_ = next_read_offset;
+  unsigned int ring_value = *(uint32_t*)&this->buffer_[read_offset];
+  return xe::byte_swap(ring_value);
+}
 }  // namespace xe
 
 #endif  // XENIA_BASE_RING_BUFFER_H_