GS: Less memcmp

It has a lot of work to do for its fancier return value

pcsx2/GS/GSExtra.h

@@ -35,6 +35,64 @@ inline std::wstring convert_utf8_to_utf16(const std::string& utf8_string)
 }
 #endif
 
+/// Like `memcmp(&a, &b, sizeof(T)) == 0` but faster
+template <typename T>
+__forceinline bool BitEqual(const T& a, const T& b)
+{
+#if _M_SSE >= 0x501
+	if (alignof(T) >= 32)
+	{
+		GSVector8i eq = GSVector8i::xffffffff();
+		for (size_t i = 0; i < sizeof(T) / 32; i++)
+			eq &= reinterpret_cast<const GSVector8i*>(&a)[i].eq8(reinterpret_cast<const GSVector8i*>(&b)[i]);
+		return eq.alltrue();
+	}
+#endif
+	GSVector4i eq = GSVector4i::xffffffff();
+	if (alignof(T) >= 16)
+	{
+		for (size_t i = 0; i < sizeof(T) / 16; i++)
+			eq &= reinterpret_cast<const GSVector4i*>(&a)[i].eq8(reinterpret_cast<const GSVector4i*>(&b)[i]);
+		return eq.alltrue();
+	}
+	const char* ac = reinterpret_cast<const char*>(&a);
+	const char* bc = reinterpret_cast<const char*>(&b);
+	size_t i = 0;
+	if (sizeof(T) >= 16)
+	{
+		for (; i < sizeof(T) - 15; i += 16)
+			eq &= GSVector4i::load<false>(ac + i).eq8(GSVector4i::load<false>(bc + i));
+	}
+	if (i + 8 <= sizeof(T))
+	{
+		eq &= GSVector4i::loadl(ac + i).eq8(GSVector4i::loadl(bc + i));
+		i += 8;
+	}
+	bool eqb = eq.alltrue();
+	if (i + 4 <= sizeof(T))
+	{
+		u32 ai, bi;
+		memcpy(&ai, ac + i, sizeof(ai));
+		memcpy(&bi, bc + i, sizeof(bi));
+		eqb = ai == bi && eqb;
+		i += 4;
+	}
+	if (i + 2 <= sizeof(T))
+	{
+		u16 as, bs;
+		memcpy(&as, ac + i, sizeof(as));
+		memcpy(&bs, bc + i, sizeof(bs));
+		eqb = as == bs && eqb;
+		i += 2;
+	}
+	if (i != sizeof(T))
+	{
+		ASSERT(i + 1 == sizeof(T));
+		eqb = ac[i] == bc[i] && eqb;
+	}
+	return eqb;
+}
+
 // _wfopen has to be used on Windows for pathnames containing non-ASCII characters.
 inline FILE* px_fopen(const std::string& filename, const std::string& mode)
 {

pcsx2/GS/Renderers/Common/GSDevice.h

@@ -20,6 +20,7 @@
 #include "GSTexture.h"
 #include "GSVertex.h"
 #include "GS/GSAlignedClass.h"
+#include "GS/GSExtra.h"
 #include "GSOsdManager.h"
 #include <array>
 #ifdef _WIN32
@@ -328,7 +329,7 @@ struct GSHWDrawConfig
 		ColorMaskSelector(): key(0xF) {}
 		ColorMaskSelector(u32 c): key(0) { wrgba = c; }
 	};
-	struct VSConstantBuffer
+	struct alignas(16) VSConstantBuffer
 	{
 		GSVector2 vertex_scale;
 		GSVector2 vertex_offset;
@@ -336,8 +337,29 @@ struct GSHWDrawConfig
 		GSVector2 texture_offset;
 		GSVector2 point_size;
 		GSVector2i max_depth;
+		VSConstantBuffer()
+		{
+			memset(this, 0, sizeof(*this));
+		}
+		VSConstantBuffer(const VSConstantBuffer& other)
+		{
+			memcpy(this, &other, sizeof(*this));
+		}
+		VSConstantBuffer& operator=(const VSConstantBuffer& other)
+		{
+			new (this) VSConstantBuffer(other);
+			return *this;
+		}
+		bool operator==(const VSConstantBuffer& other) const
+		{
+			return BitEqual(*this, other);
+		}
+		bool operator!=(const VSConstantBuffer& other) const
+		{
+			return !(*this == other);
+		}
 	};
-	struct PSConstantBuffer
+	struct alignas(16) PSConstantBuffer
 	{
 		union
 		{
@@ -396,6 +418,27 @@ struct GSHWDrawConfig
 
 		GSVector4 half_texel;
 		GSVector4 uv_min_max;
+		PSConstantBuffer()
+		{
+			memset(this, 0, sizeof(*this));
+		}
+		PSConstantBuffer(const PSConstantBuffer& other)
+		{
+			memcpy(this, &other, sizeof(*this));
+		}
+		PSConstantBuffer& operator=(const PSConstantBuffer& other)
+		{
+			new (this) PSConstantBuffer(other);
+			return *this;
+		}
+		bool operator==(const PSConstantBuffer& other) const
+		{
+			return BitEqual(*this, other);
+		}
+		bool operator!=(const PSConstantBuffer& other) const
+		{
+			return !(*this == other);
+		}
 	};
 	struct BlendState
 	{
@@ -466,10 +509,10 @@ struct GSHWDrawConfig
 	struct AlphaSecondPass
 	{
 		bool enable;
-		PSSelector ps;
-		PSConstantBuffer cb_ps;
 		ColorMaskSelector colormask;
 		DepthStencilSelector depth;
+		PSSelector ps;
+		PSConstantBuffer cb_ps;
 	} alpha_second_pass;
 };
 

pcsx2/GS/Renderers/DX11/GSDevice11.cpp

@@ -1704,7 +1704,7 @@ void GSDevice11::RenderHW(GSHWDrawConfig& config)
 
 	if (config.alpha_second_pass.enable)
 	{
-		if (0 != memcmp(&config.cb_ps, &config.alpha_second_pass.cb_ps, sizeof(config.cb_ps)))
+		if (config.cb_ps != config.alpha_second_pass.cb_ps)
 		{
 			cb_ps = convertCB(config.alpha_second_pass.cb_ps, config.alpha_second_pass.ps.atst);
 		}

pcsx2/GS/Renderers/OpenGL/GSDeviceOGL.cpp

@@ -2199,7 +2199,7 @@ void GSDeviceOGL::RenderHW(GSHWDrawConfig& config)
 
 	if (config.alpha_second_pass.enable)
 	{
-		if (0 != memcmp(&config.cb_ps, &config.alpha_second_pass.cb_ps, sizeof(config.cb_ps)))
+		if (config.cb_ps != config.alpha_second_pass.cb_ps)
 		{
 			cb_ps = convertCB(config.alpha_second_pass.cb_ps, config.alpha_second_pass.ps.atst);
 			SetupCB(&cb_vs, &cb_ps);