From 9084ef35b47ff816cc046b0559ec17d966b6d00a Mon Sep 17 00:00:00 2001
From: TellowKrinkle <tellowkrinkle@gmail.com>
Date: Mon, 16 May 2022 16:57:06 -0500
Subject: [PATCH] GS:SW: Restore zequal for optimization

---
 pcsx2/GS/Renderers/SW/GSDrawScanline.cpp      | 27 +++++++++++++++----
 .../SW/GSDrawScanlineCodeGenerator.all.cpp    | 20 +++++++++++---
 pcsx2/GS/Renderers/SW/GSRendererSW.cpp        |  1 +
 pcsx2/GS/Renderers/SW/GSScanlineEnvironment.h |  1 +
 4 files changed, 41 insertions(+), 8 deletions(-)

diff --git a/pcsx2/GS/Renderers/SW/GSDrawScanline.cpp b/pcsx2/GS/Renderers/SW/GSDrawScanline.cpp
index a0df154aa7..02fd0203ce 100644
--- a/pcsx2/GS/Renderers/SW/GSDrawScanline.cpp
+++ b/pcsx2/GS/Renderers/SW/GSDrawScanline.cpp
@@ -98,6 +98,7 @@ void GSDrawScanline::BeginDraw(const GSRasterizerData* data)
 	sel.fb = m_global.sel.fb;
 	sel.zb = m_global.sel.zb;
 	sel.zoverflow = m_global.sel.zoverflow;
+	sel.zequal = m_global.sel.zequal;
 	sel.notest = m_global.sel.notest;
 
 	m_sp = m_sp_map[sel];
@@ -161,7 +162,7 @@ void GSDrawScanline::CSetupPrim(const GSVertexSW* vertex, const u32* index, cons
 				}
 			}
 
-			if (has_z)
+			if (has_z && !sel.zequal)
 			{
 				const GSVector4 dz = GSVector4::broadcast64(&dscan.p.z);
 				const VectorF dzf(static_cast<float>(dscan.p.F64[1]));
@@ -357,8 +358,20 @@ void GSDrawScanline::CDrawScanline(int pixels, int left, int top, const GSVertex
 		if (sel.zb)
 		{
 			VectorF zbase = VectorF::broadcast64(&scan.p.z);
-			z0 = zbase.add64(VectorF::f32to64(&local.d[skip].z.F32[0]));
-			z1 = zbase.add64(VectorF::f32to64(&local.d[skip].z.F32[vlen/2]));
+			if (sel.zequal)
+			{
+#if _M_SSE >= 0x501
+				z0 = GSVector8::cast(GSVector8i::broadcast32(zbase.extract<0>().f64toi32()));
+#else
+				z0 = GSVector4::cast(zbase.f64toi32());
+				z0 = z0.upld(z0);
+#endif
+			}
+			else
+			{
+				z0 = zbase.add64(VectorF::f32to64(&local.d[skip].z.F32[0]));
+				z1 = zbase.add64(VectorF::f32to64(&local.d[skip].z.F32[vlen/2]));
+			}
 		}
 	}
 
@@ -449,7 +462,11 @@ void GSDrawScanline::CDrawScanline(int pixels, int left, int top, const GSVertex
 
 				if (sel.prim != GS_SPRITE_CLASS)
 				{
-					if (sel.zoverflow)
+					if (sel.zequal)
+					{
+						zs = VectorI::cast(z0);
+					}
+					else if (sel.zoverflow)
 					{
 						// SSE only has double to int32 conversion, no double to uint32
 						// Work around this by subtracting 0x80000000 before converting, then adding it back after
@@ -1492,7 +1509,7 @@ void GSDrawScanline::CDrawScanline(int pixels, int left, int top, const GSVertex
 
 		if (sel.prim != GS_SPRITE_CLASS)
 		{
-			if (sel.zb)
+			if (sel.zb && !sel.zequal)
 			{
 #if _M_SSE >= 0x501
 				GSVector8 add = GSVector8::broadcast64(&local.d8.p.z);
diff --git a/pcsx2/GS/Renderers/SW/GSDrawScanlineCodeGenerator.all.cpp b/pcsx2/GS/Renderers/SW/GSDrawScanlineCodeGenerator.all.cpp
index e47d93ee80..848ba98c47 100644
--- a/pcsx2/GS/Renderers/SW/GSDrawScanlineCodeGenerator.all.cpp
+++ b/pcsx2/GS/Renderers/SW/GSDrawScanlineCodeGenerator.all.cpp
@@ -720,7 +720,17 @@ void GSDrawScanlineCodeGenerator2::Init()
 				paddw(f, ptr[a1 + offsetof(GSScanlineLocalData::skip, f)]);
 			}
 
-			if (m_sel.zb)
+			if (m_sel.zb && m_sel.zequal)
+			{
+				Xmm zx(_z.getIdx());
+				cvttsd2si(rax, ptr[a3 + offsetof(GSVertexSW, p.z)]);
+				movd(zx, eax);
+				if (hasAVX2)
+					vpbroadcastd(_z, zx);
+				else
+					pshufd(_z, _z, _MM_SHUFFLE(0, 0, 0, 0));
+			}
+			else if (m_sel.zb)
 			{
 				// z = vp.zzzz() + m_local.d[skip].z;
 				broadcastsd(xym1, ptr[a3 + offsetof(GSVertexSW, p.z)]); // v.p.z
@@ -908,7 +918,7 @@ void GSDrawScanlineCodeGenerator2::Step()
 	{
 		// z += m_local.d4.z;
 
-		if (m_sel.zb)
+		if (m_sel.zb && !m_sel.zequal)
 		{
 			broadcastsd(xym7, _rip_local_d_p(z));
 			addpd(_z, xym7);
@@ -1054,7 +1064,11 @@ void GSDrawScanlineCodeGenerator2::TestZ(const XYm& temp1, const XYm& temp2)
 
 	if (m_sel.prim != GS_SPRITE_CLASS)
 	{
-		if (m_sel.zoverflow)
+		if (m_sel.zequal)
+		{
+			movdqa(xym0, _z);
+		}
+		else if (m_sel.zoverflow)
 		{
 			// GSVector4i zl = z0.add64(VectorF::m_xc1e00000000fffff).f64toi32();
 			// GSVector4i zh = z1.add64(VectorF::m_xc1e00000000fffff).f64toi32();
diff --git a/pcsx2/GS/Renderers/SW/GSRendererSW.cpp b/pcsx2/GS/Renderers/SW/GSRendererSW.cpp
index 957baaec15..9b08a5d9c4 100644
--- a/pcsx2/GS/Renderers/SW/GSRendererSW.cpp
+++ b/pcsx2/GS/Renderers/SW/GSRendererSW.cpp
@@ -1288,6 +1288,7 @@ bool GSRendererSW::GetScanlineGlobalData(SharedData* data)
 
 		gd.sel.zpsm = GSLocalMemory::m_psm[context->ZBUF.PSM].fmt;
 		gd.sel.ztst = ztest ? context->TEST.ZTST : (int)ZTST_ALWAYS;
+		gd.sel.zequal = !!m_vt.m_eq.z;
 		gd.sel.zoverflow = (u32)GSVector4i(m_vt.m_max.p).z == 0x80000000U;
 		gd.sel.zclamp = (u32)GSVector4i(m_vt.m_max.p).z > z_max;
 	}
diff --git a/pcsx2/GS/Renderers/SW/GSScanlineEnvironment.h b/pcsx2/GS/Renderers/SW/GSScanlineEnvironment.h
index f9250e48ff..a6696a6473 100644
--- a/pcsx2/GS/Renderers/SW/GSScanlineEnvironment.h
+++ b/pcsx2/GS/Renderers/SW/GSScanlineEnvironment.h
@@ -64,6 +64,7 @@ union GSScanlineSelector
 		u32 mmin   : 2; // 54
 		u32 notest : 1; // 55 (no ztest, no atest, no date, no scissor test, and horizontally aligned to 4 pixels)
 		// TODO: 1D texture flag? could save 2 texture reads and 4 lerps with bilinear, and also the texture coordinate clamp/wrap code in one direction
+		u32 zequal : 1; // 56
 		u32 breakpoint : 1; // Insert a trap to stop the program, helpful to stop debugger on a program
 	};