Add codegen for one of the triangle variants

src/CMakeLists.txt

@@ -1487,6 +1487,7 @@ set (PCH_SOURCES
 	r_compiler/fixedfunction/drawwallcodegen.cpp
 	r_compiler/fixedfunction/drawcolumncodegen.cpp
 	r_compiler/fixedfunction/drawskycodegen.cpp
+	r_compiler/fixedfunction/drawtrianglecodegen.cpp
 	r_data/sprites.cpp
 	r_data/voxels.cpp
 	r_data/renderstyle.cpp

src/r_compiler/fixedfunction/drawtrianglecodegen.cpp

@@ -0,0 +1,483 @@
+/*
+**  DrawTriangle code generation
+**  Copyright (c) 2016 Magnus Norddahl
+**
+**  This software is provided 'as-is', without any express or implied
+**  warranty.  In no event will the authors be held liable for any damages
+**  arising from the use of this software.
+**
+**  Permission is granted to anyone to use this software for any purpose,
+**  including commercial applications, and to alter it and redistribute it
+**  freely, subject to the following restrictions:
+**
+**  1. The origin of this software must not be misrepresented; you must not
+**     claim that you wrote the original software. If you use this software
+**     in a product, an acknowledgment in the product documentation would be
+**     appreciated but is not required.
+**  2. Altered source versions must be plainly marked as such, and must not be
+**     misrepresented as being the original software.
+**  3. This notice may not be removed or altered from any source distribution.
+**
+*/
+
+#include "i_system.h"
+#include "r_compiler/llvm_include.h"
+#include "r_compiler/fixedfunction/drawtrianglecodegen.h"
+#include "r_compiler/ssa/ssa_function.h"
+#include "r_compiler/ssa/ssa_scope.h"
+#include "r_compiler/ssa/ssa_for_block.h"
+#include "r_compiler/ssa/ssa_if_block.h"
+#include "r_compiler/ssa/ssa_stack.h"
+#include "r_compiler/ssa/ssa_function.h"
+#include "r_compiler/ssa/ssa_struct_type.h"
+#include "r_compiler/ssa/ssa_value.h"
+
+void DrawTriangleCodegen::Generate(TriDrawVariant variant, bool truecolor, SSAValue args, SSAValue thread_data)
+{
+	LoadArgs(variant, truecolor, args, thread_data);
+	Setup(variant, truecolor);
+	LoopBlockY(variant, truecolor);
+}
+
+void DrawTriangleCodegen::Setup(TriDrawVariant variant, bool truecolor)
+{
+	int pixelsize = truecolor ? 4 : 1;
+
+	// 28.4 fixed-point coordinates
+	Y1 = SSAInt(SSAFloat::round(16.0f * v1.y), false);
+	Y2 = SSAInt(SSAFloat::round(16.0f * v2.y), false);
+	Y3 = SSAInt(SSAFloat::round(16.0f * v3.y), false);
+
+	X1 = SSAInt(SSAFloat::round(16.0f * v1.x), false);
+	X2 = SSAInt(SSAFloat::round(16.0f * v2.x), false);
+	X3 = SSAInt(SSAFloat::round(16.0f * v3.x), false);
+
+	// Deltas
+	DX12 = X1 - X2;
+	DX23 = X2 - X3;
+	DX31 = X3 - X1;
+
+	DY12 = Y1 - Y2;
+	DY23 = Y2 - Y3;
+	DY31 = Y3 - Y1;
+
+	// Fixed-point deltas
+	FDX12 = DX12 << 4;
+	FDX23 = DX23 << 4;
+	FDX31 = DX31 << 4;
+
+	FDY12 = DY12 << 4;
+	FDY23 = DY23 << 4;
+	FDY31 = DY31 << 4;
+
+	// Bounding rectangle
+	minx = SSAInt::MAX((SSAInt::MIN(SSAInt::MIN(X1, X2), X3) + 0xF) >> 4, clipleft);
+	maxx = SSAInt::MIN((SSAInt::MAX(SSAInt::MAX(X1, X2), X3) + 0xF) >> 4, clipright - 1);
+	miny = SSAInt::MAX((SSAInt::MIN(SSAInt::MIN(Y1, Y2), Y3) + 0xF) >> 4, cliptop);
+	maxy = SSAInt::MIN((SSAInt::MAX(SSAInt::MAX(Y1, Y2), Y3) + 0xF) >> 4, clipbottom - 1);
+
+	SSAIfBlock if0;
+	if0.if_block(minx >= maxx || miny >= maxy);
+	if0.end_retvoid();
+
+	// Start in corner of 8x8 block
+	minx = minx & ~(q - 1);
+	miny = miny & ~(q - 1);
+
+	dest = dest[miny * pitch * pixelsize];
+	subsectorGBuffer = subsectorGBuffer[miny * pitch];
+
+	// Half-edge constants
+	C1 = DY12 * X1 - DX12 * Y1;
+	C2 = DY23 * X2 - DX23 * Y2;
+	C3 = DY31 * X3 - DX31 * Y3;
+
+	// Correct for fill convention
+	SSAIfBlock if1;
+	if1.if_block(DY12 < SSAInt(0) || (DY12 == SSAInt(0) && DX12 > SSAInt(0)));
+		stack_C1.store(C1 + 1);
+	if1.else_block();
+		stack_C1.store(C1);
+	if1.end_block();
+	C1 = stack_C1.load();
+	SSAIfBlock if2;
+	if2.if_block(DY23 < SSAInt(0) || (DY23 == SSAInt(0) && DX23 > SSAInt(0)));
+		stack_C2.store(C2 + 1);
+	if2.else_block();
+		stack_C2.store(C2);
+	if2.end_block();
+	C2 = stack_C2.load();
+	SSAIfBlock if3;
+	if3.if_block(DY31 < SSAInt(0) || (DY31 == SSAInt(0) && DX31 > SSAInt(0)));
+		stack_C3.store(C3 + 1);
+	if3.else_block();
+		stack_C3.store(C3);
+	if3.end_block();
+	C3 = stack_C3.load();
+
+	// Gradients
+	gradWX = gradx(v1.x, v1.y, v2.x, v2.y, v3.x, v3.y, v1.w, v2.w, v3.w);
+	gradWY = grady(v1.x, v1.y, v2.x, v2.y, v3.x, v3.y, v1.w, v2.w, v3.w);
+	startW = v1.w + gradWX * (SSAFloat(minx) - v1.x) + gradWY * (SSAFloat(miny) - v1.y);
+	for (int i = 0; i < TriVertex::NumVarying; i++)
+	{
+		gradVaryingX[i] = gradx(v1.x, v1.y, v2.x, v2.y, v3.x, v3.y, v1.varying[i] * v1.w, v2.varying[i] * v2.w, v3.varying[i] * v3.w);
+		gradVaryingY[i] = grady(v1.x, v1.y, v2.x, v2.y, v3.x, v3.y, v1.varying[i] * v1.w, v2.varying[i] * v2.w, v3.varying[i] * v3.w);
+		startVarying[i] = v1.varying[i] * v1.w + gradVaryingX[i] * (SSAFloat(minx) - v1.x) + gradVaryingY[i] * (SSAFloat(miny) - v1.y);
+	}
+}
+
+SSAFloat DrawTriangleCodegen::gradx(SSAFloat x0, SSAFloat y0, SSAFloat x1, SSAFloat y1, SSAFloat x2, SSAFloat y2, SSAFloat c0, SSAFloat c1, SSAFloat c2)
+{
+	SSAFloat top = (c1 - c2) * (y0 - y2) - (c0 - c2) * (y1 - y2);
+	SSAFloat bottom = (x1 - x2) * (y0 - y2) - (x0 - x2) * (y1 - y2);
+	return top / bottom;
+}
+
+SSAFloat DrawTriangleCodegen::grady(SSAFloat x0, SSAFloat y0, SSAFloat x1, SSAFloat y1, SSAFloat x2, SSAFloat y2, SSAFloat c0, SSAFloat c1, SSAFloat c2)
+{
+	SSAFloat top = (c1 - c2) * (x0 - x2) - (c0 - c2) * (x1 - x2);
+	SSAFloat bottom = (x0 - x2) * (y1 - y2) - (x1 - x2) * (y0 - y2);
+	return top / bottom;
+}
+
+void DrawTriangleCodegen::LoopBlockY(TriDrawVariant variant, bool truecolor)
+{
+	int pixelsize = truecolor ? 4 : 1;
+
+	stack_y.store(miny);
+	stack_dest.store(dest);
+	stack_subsectorGBuffer.store(subsectorGBuffer);
+
+	SSAForBlock loop;
+	y = stack_y.load();
+	dest = stack_dest.load();
+	subsectorGBuffer = stack_subsectorGBuffer.load();
+	loop.loop_block(y < maxy, 0);
+	{
+		SSAIfBlock branch;
+		branch.if_block((y / q) % thread.num_cores == thread.core);
+		{
+			LoopBlockX(variant, truecolor);
+		}
+		branch.end_block();
+
+		stack_dest.store(dest[q * pitch * pixelsize]);
+		stack_subsectorGBuffer.store(subsectorGBuffer[q * pitch]);
+		stack_y.store(y + q);
+	}
+	loop.end_block();
+}
+
+void DrawTriangleCodegen::LoopBlockX(TriDrawVariant variant, bool truecolor)
+{
+	stack_x.store(minx);
+
+	SSAForBlock loop;
+	x = stack_x.load();
+	loop.loop_block(x < maxx, 0);
+	{
+		// Corners of block
+		x0 = x << 4;
+		x1 = (x + q - 1) << 4;
+		y0 = y << 4;
+		y1 = (y + q - 1) << 4;
+
+		// Evaluate half-space functions
+		SSABool a00 = C1 + DX12 * y0 - DY12 * x0 > SSAInt(0);
+		SSABool a10 = C1 + DX12 * y0 - DY12 * x1 > SSAInt(0);
+		SSABool a01 = C1 + DX12 * y1 - DY12 * x0 > SSAInt(0);
+		SSABool a11 = C1 + DX12 * y1 - DY12 * x1 > SSAInt(0);
+		
+		SSAInt a = (a00.zext_int() << 0) | (a10.zext_int() << 1) | (a01.zext_int() << 2) | (a11.zext_int() << 3);
+
+		SSABool b00 = C2 + DX23 * y0 - DY23 * x0 > SSAInt(0);
+		SSABool b10 = C2 + DX23 * y0 - DY23 * x1 > SSAInt(0);
+		SSABool b01 = C2 + DX23 * y1 - DY23 * x0 > SSAInt(0);
+		SSABool b11 = C2 + DX23 * y1 - DY23 * x1 > SSAInt(0);
+		SSAInt b = (b00.zext_int() << 0) | (b10.zext_int() << 1) | (b01.zext_int() << 2) | (b11.zext_int() << 3);
+
+		SSABool c00 = C3 + DX31 * y0 - DY31 * x0 > SSAInt(0);
+		SSABool c10 = C3 + DX31 * y0 - DY31 * x1 > SSAInt(0);
+		SSABool c01 = C3 + DX31 * y1 - DY31 * x0 > SSAInt(0);
+		SSABool c11 = C3 + DX31 * y1 - DY31 * x1 > SSAInt(0);
+		SSAInt c = (c00.zext_int() << 0) | (c10.zext_int() << 1) | (c01.zext_int() << 2) | (c11.zext_int() << 3);
+		
+		// Skip block when outside an edge
+		SSAIfBlock branch;
+		branch.if_block(!(a == SSAInt(0) || b == SSAInt(0) || c == SSAInt(0)));
+
+		// Check if block needs clipping
+		SSABool clipneeded = clipleft > x || clipright < (x + q) || cliptop > y || clipbottom < (y + q);
+
+		// Calculate varying variables for affine block
+		SSAFloat offx0 = SSAFloat(x - minx) + 0.5f;
+		SSAFloat offy0 = SSAFloat(y - miny) + 0.5f;
+		SSAFloat offx1 = offx0 + SSAFloat(q);
+		SSAFloat offy1 = offy0 + SSAFloat(q);
+		SSAFloat rcpWTL = 1.0f / (startW + offx0 * gradWX + offy0 * gradWY);
+		SSAFloat rcpWTR = 1.0f / (startW + offx1 * gradWX + offy0 * gradWY);
+		SSAFloat rcpWBL = 1.0f / (startW + offx0 * gradWX + offy1 * gradWY);
+		SSAFloat rcpWBR = 1.0f / (startW + offx1 * gradWX + offy1 * gradWY);
+		for (int i = 0; i < TriVertex::NumVarying; i++)
+		{
+			varyingTL[i] = (startVarying[i] + offx0 * gradVaryingX[i] + offy0 * gradVaryingY[i]) * rcpWTL;
+			varyingTR[i] = (startVarying[i] + offx1 * gradVaryingX[i] + offy0 * gradVaryingY[i]) * rcpWTR;
+			varyingBL[i] = ((startVarying[i] + offx0 * gradVaryingX[i] + offy1 * gradVaryingY[i]) * rcpWBL - varyingTL[i]) * (1.0f / q);
+			varyingBR[i] = ((startVarying[i] + offx1 * gradVaryingX[i] + offy1 * gradVaryingY[i]) * rcpWBR - varyingTR[i]) * (1.0f / q);
+		}
+
+		SSAFloat globVis = SSAFloat(1706.0f);
+		SSAFloat vis = globVis / rcpWTL;
+		SSAFloat shade = 64.0f - (SSAFloat(light * 255 / 256) + 12.0f) * 32.0f / 128.0f;
+		SSAFloat lightscale = SSAFloat::clamp((shade - SSAFloat::MIN(SSAFloat(24.0f), vis)) / 32.0f, SSAFloat(0.0f), SSAFloat(31.0f / 32.0f));
+		diminishedlight = SSAInt(SSAFloat::clamp((1.0f - lightscale) * 256.0f + 0.5f, SSAFloat(0.0f), SSAFloat(256.0f)), false);
+
+		// Accept whole block when totally covered
+		SSAIfBlock branch_covered;
+		branch_covered.if_block(a == SSAInt(0xF) && b == SSAInt(0xF) && c == SSAInt(0xF) && !clipneeded);
+		{
+			LoopFullBlock(variant, truecolor);
+		}
+		branch_covered.else_block();
+		{
+			LoopPartialBlock(variant, truecolor);
+		}
+		branch_covered.end_block();
+
+		branch.end_block();
+
+		stack_x.store(x + q);
+	}
+	loop.end_block();
+}
+
+void DrawTriangleCodegen::LoopFullBlock(TriDrawVariant variant, bool truecolor)
+{
+	int pixelsize = truecolor ? 4 : 1;
+
+	stack_iy.store(SSAInt(0));
+	stack_buffer.store(dest);
+	stack_subsectorbuffer.store(subsectorGBuffer);
+
+	SSAForBlock loopy;
+	SSAInt iy = stack_iy.load();
+	SSAUBytePtr buffer = stack_buffer.load();
+	SSAIntPtr subsectorbuffer = stack_subsectorbuffer.load();
+	loopy.loop_block(iy < SSAInt(q), q);
+	{
+		SSAInt varyingStep[TriVertex::NumVarying];
+		for (int i = 0; i < TriVertex::NumVarying; i++)
+		{
+			SSAFloat pos = varyingTL[i] + varyingBL[i] * SSAFloat(iy);
+			SSAFloat step = (varyingTR[i] + varyingBR[i] * SSAFloat(iy) - pos) * (1.0f / q);
+
+			stack_varying[i].store(SSAInt((pos - SSAFloat::floor(pos)) * SSAFloat((float)0x100000000LL), true));
+			varyingStep[i] = SSAInt(step * SSAFloat((float)0x100000000LL), true);
+		}
+
+		stack_ix.store(x);
+		SSAForBlock loopx;
+		SSAInt ix = stack_ix.load();
+		SSAInt varying[TriVertex::NumVarying];
+		for (int i = 0; i < TriVertex::NumVarying; i++)
+			varying[i] = stack_varying[i].load();
+		loopx.loop_block(ix < x + q, q);
+		{
+			SSAIfBlock branch;
+			branch.if_block(subsectorbuffer[ix].load(true) >= subsectorDepth);
+			{
+				if (truecolor)
+					ProcessPixel(buffer[ix * 4], subsectorbuffer[ix], varying, variant, truecolor);
+				else
+					ProcessPixel(buffer[ix], subsectorbuffer[ix], varying, variant, truecolor);
+			}
+			branch.end_block();
+
+			for (int i = 0; i < TriVertex::NumVarying; i++)
+				stack_varying[i].store(varying[i] + varyingStep[i]);
+
+			stack_ix.store(ix + 1);
+		}
+		loopx.end_block();
+
+		stack_buffer.store(buffer[pitch * pixelsize]);
+		stack_subsectorbuffer.store(subsectorbuffer[pitch]);
+		stack_iy.store(iy + 1);
+	}
+	loopy.end_block();
+}
+
+void DrawTriangleCodegen::LoopPartialBlock(TriDrawVariant variant, bool truecolor)
+{
+	int pixelsize = truecolor ? 4 : 1;
+
+	stack_CY1.store(C1 + DX12 * y0 - DY12 * x0);
+	stack_CY2.store(C2 + DX23 * y0 - DY23 * x0);
+	stack_CY3.store(C3 + DX31 * y0 - DY31 * x0);
+	stack_iy.store(SSAInt(0));
+	stack_buffer.store(dest);
+	stack_subsectorbuffer.store(subsectorGBuffer);
+
+	SSAForBlock loopy;
+	SSAInt iy = stack_iy.load();
+	SSAUBytePtr buffer = stack_buffer.load();
+	SSAIntPtr subsectorbuffer = stack_subsectorbuffer.load();
+	SSAInt CY1 = stack_CY1.load();
+	SSAInt CY2 = stack_CY2.load();
+	SSAInt CY3 = stack_CY3.load();
+	loopy.loop_block(iy < SSAInt(q), q);
+	{
+		SSAInt varyingStep[TriVertex::NumVarying];
+		for (int i = 0; i < TriVertex::NumVarying; i++)
+		{
+			SSAFloat pos = varyingTL[i] + varyingBL[i] * SSAFloat(iy);
+			SSAFloat step = (varyingTR[i] + varyingBR[i] * SSAFloat(iy) - pos) * (1.0f / q);
+
+			stack_varying[i].store(SSAInt((pos - SSAFloat::floor(pos)) * SSAFloat((float)0x100000000LL), true));
+			varyingStep[i] = SSAInt(step * SSAFloat((float)0x100000000LL), true);
+		}
+
+		stack_CX1.store(CY1);
+		stack_CX2.store(CY2);
+		stack_CX3.store(CY3);
+		stack_ix.store(SSAInt(0));
+
+		SSAForBlock loopx;
+		SSAInt ix = stack_ix.load();
+		SSAInt CX1 = stack_CX1.load();
+		SSAInt CX2 = stack_CX2.load();
+		SSAInt CX3 = stack_CX3.load();
+		SSAInt varying[TriVertex::NumVarying];
+		for (int i = 0; i < TriVertex::NumVarying; i++)
+			varying[i] = stack_varying[i].load();
+		loopx.loop_block(ix < SSAInt(q), q);
+		{
+			SSABool visible = (ix + x >= clipleft) && (ix + x < clipright) && (cliptop <= y + iy) && (clipbottom > y + iy);
+
+			SSAIfBlock branch;
+			branch.if_block(CX1 > SSAInt(0) && CX2 > SSAInt(0) && CX3 > SSAInt(0) && visible && subsectorbuffer[ix + x].load(true) >= subsectorDepth);
+			{
+				if (truecolor)
+					ProcessPixel(buffer[(ix + x) * 4], subsectorbuffer[ix + x], varying, variant, truecolor);
+				else
+					ProcessPixel(buffer[ix + x], subsectorbuffer[ix + x], varying, variant, truecolor);
+			}
+			branch.end_block();
+
+			for (int i = 0; i < TriVertex::NumVarying; i++)
+				stack_varying[i].store(varying[i] + varyingStep[i]);
+
+			stack_CX1.store(CX1 - FDY12);
+			stack_CX2.store(CX2 - FDY23);
+			stack_CX3.store(CX3 - FDY31);
+			stack_ix.store(ix + 1);
+		}
+		loopx.end_block();
+
+		stack_CY1.store(CY1 + FDX12);
+		stack_CY2.store(CY2 + FDX23);
+		stack_CY3.store(CY3 + FDX31);
+		stack_buffer.store(buffer[pitch * pixelsize]);
+		stack_subsectorbuffer.store(subsectorbuffer[pitch]);
+		stack_iy.store(iy + 1);
+	}
+	loopy.end_block();
+}
+
+void DrawTriangleCodegen::ProcessPixel(SSAUBytePtr buffer, SSAIntPtr subsectorbuffer, SSAInt *varying, TriDrawVariant variant, bool truecolor)
+{
+	SSAInt ufrac = varying[0];
+	SSAInt vfrac = varying[1];
+
+	SSAInt upos = ((ufrac >> 16) * textureWidth) >> 16;
+	SSAInt vpos = ((vfrac >> 16) * textureHeight) >> 16;
+	SSAInt uvoffset = upos * textureHeight + vpos;
+
+	if (truecolor)
+	{
+		SSAVec4i fg = texturePixels[uvoffset * 4].load_vec4ub(true);
+		SSAInt fg_alpha = fg[3];
+		fg = (fg * diminishedlight) >> 8;
+		fg.insert(3, fg_alpha);
+
+		SSAIfBlock branch_transparency;
+		branch_transparency.if_block(fg_alpha > SSAInt(127));
+		{
+			buffer.store_vec4ub(fg);
+		}
+		branch_transparency.end_block();
+	}
+	else
+	{
+		SSAUByte palindex = texturePixels[uvoffset].load(true);
+		SSAIfBlock branch_transparency;
+		branch_transparency.if_block(!(palindex.zext_int() == SSAInt(0)));
+		{
+			buffer.store(palindex);
+		}
+		branch_transparency.end_block();
+	}
+}
+
+void DrawTriangleCodegen::LoadArgs(TriDrawVariant variant, bool truecolor, SSAValue args, SSAValue thread_data)
+{
+	dest = args[0][0].load(true);
+	pitch = args[0][1].load(true);
+	v1 = LoadTriVertex(args[0][2].load(true));
+	v2 = LoadTriVertex(args[0][3].load(true));
+	v3 = LoadTriVertex(args[0][4].load(true));
+	clipleft = args[0][5].load(true);
+	clipright = args[0][6].load(true);
+	cliptop = args[0][7].load(true);
+	clipbottom = args[0][8].load(true);
+	texturePixels = args[0][9].load(true);
+	textureWidth = args[0][10].load(true);
+	textureHeight = args[0][11].load(true);
+	solidcolor = args[0][12].load(true);
+	LoadUniforms(args[0][13].load(true));
+	stencilValues = args[0][14].load(true);
+	stencilMasks = args[0][15].load(true);
+	stencilPitch = args[0][16].load(true);
+	stencilTestValue = args[0][17].load(true);
+	stencilWriteValue = args[0][18].load(true);
+	subsectorGBuffer = args[0][19].load(true);
+
+	thread.core = thread_data[0][0].load(true);
+	thread.num_cores = thread_data[0][1].load(true);
+}
+
+SSATriVertex DrawTriangleCodegen::LoadTriVertex(SSAValue ptr)
+{
+	SSATriVertex v;
+	v.x = ptr[0][0].load(true);
+	v.y = ptr[0][1].load(true);
+	v.z = ptr[0][2].load(true);
+	v.w = ptr[0][3].load(true);
+	for (int i = 0; i < TriVertex::NumVarying; i++)
+		v.varying[i] = ptr[0][4 + i].load(true);
+	return v;
+}
+
+void DrawTriangleCodegen::LoadUniforms(SSAValue uniforms)
+{
+	light = uniforms[0][0].load(true);
+	subsectorDepth = uniforms[0][1].load(true);
+
+	SSAShort light_alpha = uniforms[0][2].load(true);
+	SSAShort light_red = uniforms[0][3].load(true);
+	SSAShort light_green = uniforms[0][4].load(true);
+	SSAShort light_blue = uniforms[0][5].load(true);
+	SSAShort fade_alpha = uniforms[0][6].load(true);
+	SSAShort fade_red = uniforms[0][7].load(true);
+	SSAShort fade_green = uniforms[0][8].load(true);
+	SSAShort fade_blue = uniforms[0][9].load(true);
+	SSAShort desaturate = uniforms[0][10].load(true);
+	SSAInt flags = uniforms[0][11].load(true);
+	shade_constants.light = SSAVec4i(light_blue.zext_int(), light_green.zext_int(), light_red.zext_int(), light_alpha.zext_int());
+	shade_constants.fade = SSAVec4i(fade_blue.zext_int(), fade_green.zext_int(), fade_red.zext_int(), fade_alpha.zext_int());
+	shade_constants.desaturate = desaturate.zext_int();
+
+	is_simple_shade = (flags & TriUniforms::simple_shade) == SSAInt(TriUniforms::simple_shade);
+	is_nearest_filter = (flags & TriUniforms::nearest_filter) == SSAInt(TriUniforms::nearest_filter);
+}

src/r_compiler/fixedfunction/drawtrianglecodegen.h

@@ -0,0 +1,115 @@
+/*
+**  DrawTriangle code generation
+**  Copyright (c) 2016 Magnus Norddahl
+**
+**  This software is provided 'as-is', without any express or implied
+**  warranty.  In no event will the authors be held liable for any damages
+**  arising from the use of this software.
+**
+**  Permission is granted to anyone to use this software for any purpose,
+**  including commercial applications, and to alter it and redistribute it
+**  freely, subject to the following restrictions:
+**
+**  1. The origin of this software must not be misrepresented; you must not
+**     claim that you wrote the original software. If you use this software
+**     in a product, an acknowledgment in the product documentation would be
+**     appreciated but is not required.
+**  2. Altered source versions must be plainly marked as such, and must not be
+**     misrepresented as being the original software.
+**  3. This notice may not be removed or altered from any source distribution.
+**
+*/
+
+#pragma once
+
+#include "drawercodegen.h"
+
+struct SSATriVertex
+{
+	SSAFloat x, y, z, w;
+	SSAFloat varying[TriVertex::NumVarying];
+};
+
+class DrawTriangleCodegen : public DrawerCodegen
+{
+public:
+	void Generate(TriDrawVariant variant, bool truecolor, SSAValue args, SSAValue thread_data);
+
+private:
+	void LoadArgs(TriDrawVariant variant, bool truecolor, SSAValue args, SSAValue thread_data);
+	SSATriVertex LoadTriVertex(SSAValue v);
+	void LoadUniforms(SSAValue uniforms);
+	void Setup(TriDrawVariant variant, bool truecolor);
+	void LoopBlockY(TriDrawVariant variant, bool truecolor);
+	void LoopBlockX(TriDrawVariant variant, bool truecolor);
+	void LoopFullBlock(TriDrawVariant variant, bool truecolor);
+	void LoopPartialBlock(TriDrawVariant variant, bool truecolor);
+
+	void ProcessPixel(SSAUBytePtr buffer, SSAIntPtr subsectorbuffer, SSAInt *varying, TriDrawVariant variant, bool truecolor);
+
+	SSAFloat gradx(SSAFloat x0, SSAFloat y0, SSAFloat x1, SSAFloat y1, SSAFloat x2, SSAFloat y2, SSAFloat c0, SSAFloat c1, SSAFloat c2);
+	SSAFloat grady(SSAFloat x0, SSAFloat y0, SSAFloat x1, SSAFloat y1, SSAFloat x2, SSAFloat y2, SSAFloat c0, SSAFloat c1, SSAFloat c2);
+
+	SSAStack<SSAInt> stack_C1, stack_C2, stack_C3;
+	SSAStack<SSAInt> stack_y;
+	SSAStack<SSAUBytePtr> stack_dest;
+	SSAStack<SSAIntPtr> stack_subsectorGBuffer;
+	SSAStack<SSAInt> stack_x;
+	SSAStack<SSAUBytePtr> stack_buffer;
+	SSAStack<SSAIntPtr> stack_subsectorbuffer;
+	SSAStack<SSAInt> stack_iy, stack_ix;
+	SSAStack<SSAInt> stack_varying[TriVertex::NumVarying];
+	SSAStack<SSAInt> stack_CY1, stack_CY2, stack_CY3;
+	SSAStack<SSAInt> stack_CX1, stack_CX2, stack_CX3;
+
+	SSAUBytePtr dest;
+	SSAInt pitch;
+	SSATriVertex v1;
+	SSATriVertex v2;
+	SSATriVertex v3;
+	SSAInt clipleft;
+	SSAInt clipright;
+	SSAInt cliptop;
+	SSAInt clipbottom;
+	SSAUBytePtr texturePixels;
+	SSAInt textureWidth;
+	SSAInt textureHeight;
+	SSAInt solidcolor;
+
+	SSAInt light;
+	SSAInt subsectorDepth;
+	SSAShadeConstants shade_constants;
+	SSABool is_simple_shade;
+	SSABool is_nearest_filter;
+
+	SSAUBytePtr stencilValues;
+	SSAIntPtr stencilMasks;
+	SSAInt stencilPitch;
+	SSAUByte stencilTestValue;
+	SSAUByte stencilWriteValue;
+	SSAIntPtr subsectorGBuffer;
+
+	SSAWorkerThread thread;
+
+	// Block size, standard 8x8 (must be power of two)
+	const int q = 8;
+
+	SSAInt Y1, Y2, Y3;
+	SSAInt X1, X2, X3;
+	SSAInt DX12, DX23, DX31;
+	SSAInt DY12, DY23, DY31;
+	SSAInt FDX12, FDX23, FDX31;
+	SSAInt FDY12, FDY23, FDY31;
+	SSAInt minx, maxx, miny, maxy;
+	SSAInt C1, C2, C3;
+	SSAFloat gradWX, gradWY, startW;
+	SSAFloat gradVaryingX[TriVertex::NumVarying], gradVaryingY[TriVertex::NumVarying], startVarying[TriVertex::NumVarying];
+
+	SSAInt x, y;
+	SSAInt x0, x1, y0, y1;
+	SSAInt diminishedlight;
+	SSAFloat varyingTL[TriVertex::NumVarying];
+	SSAFloat varyingTR[TriVertex::NumVarying];
+	SSAFloat varyingBL[TriVertex::NumVarying];
+	SSAFloat varyingBR[TriVertex::NumVarying];
+};

src/r_compiler/llvmdrawers.cpp

@@ -26,6 +26,7 @@
 #include "r_compiler/fixedfunction/drawwallcodegen.h"
 #include "r_compiler/fixedfunction/drawcolumncodegen.h"
 #include "r_compiler/fixedfunction/drawskycodegen.h"
+#include "r_compiler/fixedfunction/drawtrianglecodegen.h"
 #include "r_compiler/ssa/ssa_function.h"
 #include "r_compiler/ssa/ssa_scope.h"
 #include "r_compiler/ssa/ssa_for_block.h"
@@ -81,6 +82,7 @@ private:
 	void CodegenDrawSpan(const char *name, DrawSpanVariant variant);
 	void CodegenDrawWall(const char *name, DrawWallVariant variant, int columns);
 	void CodegenDrawSky(const char *name, DrawSkyVariant variant, int columns);
+	void CodegenDrawTriangle(const char *name, TriDrawVariant variant, bool truecolor);
 
 	static llvm::Type *GetDrawColumnArgsStruct(llvm::LLVMContext &context);
 	static llvm::Type *GetDrawSpanArgsStruct(llvm::LLVMContext &context);
@@ -184,6 +186,13 @@ LLVMDrawersImpl::LLVMDrawersImpl()
 	CodegenDrawSky("DrawSky4", DrawSkyVariant::Single, 4);
 	CodegenDrawSky("DrawDoubleSky1", DrawSkyVariant::Double, 1);
 	CodegenDrawSky("DrawDoubleSky4", DrawSkyVariant::Double, 4);
+	CodegenDrawTriangle("TriDraw8", TriDrawVariant::Draw, false);
+	CodegenDrawTriangle("TriDraw32", TriDrawVariant::Draw, true);
+	CodegenDrawTriangle("TriDrawSubsector8", TriDrawVariant::DrawSubsector, false);
+	CodegenDrawTriangle("TriDrawSubsector32", TriDrawVariant::DrawSubsector, true);
+	CodegenDrawTriangle("TriFill8", TriDrawVariant::Fill, false);
+	CodegenDrawTriangle("TriFill32", TriDrawVariant::Fill, true);
+	CodegenDrawTriangle("TriStencil", TriDrawVariant::Stencil, false);
 
 	mProgram.CreateEE();
 
@@ -249,6 +258,13 @@ LLVMDrawersImpl::LLVMDrawersImpl()
 	DrawSky4 = mProgram.GetProcAddress<void(const DrawSkyArgs *, const WorkerThreadData *)>("DrawSky4");
 	DrawDoubleSky1 = mProgram.GetProcAddress<void(const DrawSkyArgs *, const WorkerThreadData *)>("DrawDoubleSky1");
 	DrawDoubleSky4 = mProgram.GetProcAddress<void(const DrawSkyArgs *, const WorkerThreadData *)>("DrawDoubleSky4");
+	TriDraw8 = mProgram.GetProcAddress<void(const TriDrawTriangleArgs *, WorkerThreadData *)>("TriDraw8");
+	TriDraw32 = mProgram.GetProcAddress<void(const TriDrawTriangleArgs *, WorkerThreadData *)>("TriDraw32");
+	TriDrawSubsector8 = mProgram.GetProcAddress<void(const TriDrawTriangleArgs *, WorkerThreadData *)>("TriDrawSubsector8");
+	TriDrawSubsector32 = mProgram.GetProcAddress<void(const TriDrawTriangleArgs *, WorkerThreadData *)>("TriDrawSubsector32");
+	TriFill8 = mProgram.GetProcAddress<void(const TriDrawTriangleArgs *, WorkerThreadData *)>("TriFill8");
+	TriFill32 = mProgram.GetProcAddress<void(const TriDrawTriangleArgs *, WorkerThreadData *)>("TriFill32");
+	TriStencil = mProgram.GetProcAddress<void(const TriDrawTriangleArgs *, WorkerThreadData *)>("TriStencil");
 
 #if 0
 	std::vector<uint32_t> foo(1024 * 4);
@@ -359,6 +375,25 @@ void LLVMDrawersImpl::CodegenDrawSky(const char *name, DrawSkyVariant variant, i
 		I_FatalError("verifyFunction failed for CodegenDrawSky()");
 }
 
+void LLVMDrawersImpl::CodegenDrawTriangle(const char *name, TriDrawVariant variant, bool truecolor)
+{
+	llvm::IRBuilder<> builder(mProgram.context());
+	SSAScope ssa_scope(&mProgram.context(), mProgram.module(), &builder);
+
+	SSAFunction function(name);
+	function.add_parameter(GetTriDrawTriangleArgs(mProgram.context()));
+	function.add_parameter(GetWorkerThreadDataStruct(mProgram.context()));
+	function.create_public();
+
+	DrawTriangleCodegen codegen;
+	codegen.Generate(variant, truecolor, function.parameter(0), function.parameter(1));
+
+	builder.CreateRetVoid();
+
+	if (llvm::verifyFunction(*function.func))
+		I_FatalError("verifyFunction failed for CodegenDrawTriangle()");
+}
+
 llvm::Type *LLVMDrawersImpl::GetDrawColumnArgsStruct(llvm::LLVMContext &context)
 {
 	std::vector<llvm::Type *> elements;
@@ -468,7 +503,7 @@ llvm::Type *LLVMDrawersImpl::GetWorkerThreadDataStruct(llvm::LLVMContext &contex
 llvm::Type *LLVMDrawersImpl::GetTriVertexStruct(llvm::LLVMContext &context)
 {
 	std::vector<llvm::Type *> elements;
-	for (int i = 0; i < 6; i++)
+	for (int i = 0; i < 4 + TriVertex::NumVarying; i++)
 		elements.push_back(llvm::Type::getFloatTy(context));
 	return llvm::StructType::create(context, elements, "TriVertex", false)->getPointerTo();
 }
@@ -505,9 +540,9 @@ llvm::Type *LLVMDrawersImpl::GetTriDrawTriangleArgs(llvm::LLVMContext &context)
 	std::vector<llvm::Type *> elements;
 	elements.push_back(llvm::Type::getInt8PtrTy(context));  // uint8_t *dest;
 	elements.push_back(llvm::Type::getInt32Ty(context));    // int32_t pitch;
-	elements.push_back(GetTriVertexStruct(context)->getPointerTo()); // TriVertex *v1;
+	elements.push_back(GetTriVertexStruct(context)); // TriVertex *v1;
-	elements.push_back(GetTriVertexStruct(context)->getPointerTo()); // TriVertex *v2;
+	elements.push_back(GetTriVertexStruct(context)); // TriVertex *v2;
-	elements.push_back(GetTriVertexStruct(context)->getPointerTo()); // TriVertex *v3;
+	elements.push_back(GetTriVertexStruct(context)); // TriVertex *v3;
 	elements.push_back(llvm::Type::getInt32Ty(context));    // int32_t clipleft;
 	elements.push_back(llvm::Type::getInt32Ty(context));    // int32_t clipright;
 	elements.push_back(llvm::Type::getInt32Ty(context));    // int32_t cliptop;
@@ -516,7 +551,7 @@ llvm::Type *LLVMDrawersImpl::GetTriDrawTriangleArgs(llvm::LLVMContext &context)
 	elements.push_back(llvm::Type::getInt32Ty(context));    // uint32_t textureWidth;
 	elements.push_back(llvm::Type::getInt32Ty(context));    // uint32_t textureHeight;
 	elements.push_back(llvm::Type::getInt32Ty(context));    // uint32_t solidcolor;
-	elements.push_back(GetTriUniformsStruct(context)->getPointerTo()); // const TriUniforms *uniforms;
+	elements.push_back(GetTriUniformsStruct(context)); // const TriUniforms *uniforms;
 	elements.push_back(llvm::Type::getInt8PtrTy(context));  // uint8_t *stencilValues;
 	elements.push_back(llvm::Type::getInt32PtrTy(context)); // uint32_t *stencilMasks;
 	elements.push_back(llvm::Type::getInt32Ty(context));    // int32_t stencilPitch;

src/r_compiler/llvmdrawers.h

@@ -260,6 +260,14 @@ struct TriDrawTriangleArgs
 	uint32_t *subsectorGBuffer;
 };
 
+enum class TriDrawVariant
+{
+	Draw,
+	Fill,
+	DrawSubsector,
+	Stencil,
+};
+
 class LLVMDrawers
 {
 public:
@@ -335,6 +343,14 @@ public:
 	void(*DrawDoubleSky1)(const DrawSkyArgs *, const WorkerThreadData *) = nullptr;
 	void(*DrawDoubleSky4)(const DrawSkyArgs *, const WorkerThreadData *) = nullptr;
 
+	void(*TriDraw8)(const TriDrawTriangleArgs *, WorkerThreadData *) = nullptr;
+	void(*TriDraw32)(const TriDrawTriangleArgs *, WorkerThreadData *) = nullptr;
+	void(*TriDrawSubsector8)(const TriDrawTriangleArgs *, WorkerThreadData *) = nullptr;
+	void(*TriDrawSubsector32)(const TriDrawTriangleArgs *, WorkerThreadData *) = nullptr;
+	void(*TriFill8)(const TriDrawTriangleArgs *, WorkerThreadData *) = nullptr;
+	void(*TriFill32)(const TriDrawTriangleArgs *, WorkerThreadData *) = nullptr;
+	void(*TriStencil)(const TriDrawTriangleArgs *, WorkerThreadData *) = nullptr;
+
 private:
 	static LLVMDrawers *Singleton;
 };

src/r_compiler/ssa/ssa_bool.cpp

@@ -44,6 +44,11 @@ llvm::Type *SSABool::llvm_type()
 	return llvm::Type::getInt1Ty(SSAScope::context());
 }
 
+SSAInt SSABool::zext_int()
+{
+	return SSAInt::from_llvm(SSAScope::builder().CreateZExt(v, SSAInt::llvm_type(), SSAScope::hint()));
+}
+
 SSABool operator&&(const SSABool &a, const SSABool &b)
 {
 	return SSABool::from_llvm(SSAScope::builder().CreateAnd(a.v, b.v, SSAScope::hint()));

src/r_compiler/ssa/ssa_bool.h

@@ -37,6 +37,8 @@ public:
 	static SSABool from_llvm(llvm::Value *v) { return SSABool(v); }
 	static llvm::Type *llvm_type();
 
+	SSAInt zext_int();
+
 	llvm::Value *v;
 };
 

src/r_compiler/ssa/ssa_float.cpp

@@ -24,6 +24,7 @@
 #include "ssa_float.h"
 #include "ssa_int.h"
 #include "ssa_scope.h"
+#include "ssa_bool.h"
 
 SSAFloat::SSAFloat()
 : v(0)
@@ -111,6 +112,35 @@ SSAFloat SSAFloat::fma(SSAFloat a, SSAFloat b, SSAFloat c)
 	return SSAFloat::from_llvm(SSAScope::builder().CreateCall(SSAScope::intrinsic(llvm::Intrinsic::fma, params), args, SSAScope::hint()));
 }
 
+SSAFloat SSAFloat::round(SSAFloat val)
+{
+	std::vector<llvm::Type *> params;
+	params.push_back(SSAFloat::llvm_type());
+	return SSAFloat::from_llvm(SSAScope::builder().CreateCall(SSAScope::intrinsic(llvm::Intrinsic::round, params), val.v, SSAScope::hint()));
+}
+
+SSAFloat SSAFloat::floor(SSAFloat val)
+{
+	std::vector<llvm::Type *> params;
+	params.push_back(SSAFloat::llvm_type());
+	return SSAFloat::from_llvm(SSAScope::builder().CreateCall(SSAScope::intrinsic(llvm::Intrinsic::floor, params), val.v, SSAScope::hint()));
+}
+
+SSAFloat SSAFloat::MIN(SSAFloat a, SSAFloat b)
+{
+	return SSAFloat::from_llvm(SSAScope::builder().CreateSelect((a < b).v, a.v, b.v, SSAScope::hint()));
+}
+
+SSAFloat SSAFloat::MAX(SSAFloat a, SSAFloat b)
+{
+	return SSAFloat::from_llvm(SSAScope::builder().CreateSelect((a > b).v, a.v, b.v, SSAScope::hint()));
+}
+
+SSAFloat SSAFloat::clamp(SSAFloat a, SSAFloat b, SSAFloat c)
+{
+	return SSAFloat::MAX(SSAFloat::MIN(a, c), b);
+}
+
 SSAFloat operator+(const SSAFloat &a, const SSAFloat &b)
 {
 	return SSAFloat::from_llvm(SSAScope::builder().CreateFAdd(a.v, b.v, SSAScope::hint()));

src/r_compiler/ssa/ssa_float.h

@@ -43,6 +43,11 @@ public:
 	static SSAFloat exp(SSAFloat val);
 	static SSAFloat log(SSAFloat val);
 	static SSAFloat fma(SSAFloat a, SSAFloat b, SSAFloat c);
+	static SSAFloat round(SSAFloat val);
+	static SSAFloat floor(SSAFloat val);
+	static SSAFloat MIN(SSAFloat a, SSAFloat b);
+	static SSAFloat MAX(SSAFloat a, SSAFloat b);
+	static SSAFloat clamp(SSAFloat a, SSAFloat b, SSAFloat c);
 
 	llvm::Value *v;
 };

src/r_compiler/ssa/ssa_if_block.cpp

@@ -50,3 +50,9 @@ void SSAIfBlock::end_block()
 	SSAScope::builder().CreateBr(end_basic_block);
 	SSAScope::builder().SetInsertPoint(end_basic_block);
 }
+
+void SSAIfBlock::end_retvoid()
+{
+	SSAScope::builder().CreateRetVoid();
+	SSAScope::builder().SetInsertPoint(end_basic_block);
+}

src/r_compiler/ssa/ssa_if_block.h

@@ -32,6 +32,7 @@ public:
 	void if_block(SSABool true_condition);
 	void else_block();
 	void end_block();
+	void end_retvoid();
 
 private:
 	llvm::BasicBlock *if_basic_block;

src/r_compiler/ssa/ssa_int.cpp

@@ -37,10 +37,13 @@ SSAInt::SSAInt(int constant)
 	v = llvm::ConstantInt::get(SSAScope::context(), llvm::APInt(32, constant, true));
 }
 
-SSAInt::SSAInt(SSAFloat f)
+SSAInt::SSAInt(SSAFloat f, bool uint)
 : v(0)
 {
-	v = SSAScope::builder().CreateFPToSI(f.v, llvm::Type::getInt32Ty(SSAScope::context()), SSAScope::hint());
+	if (uint)
+		v = SSAScope::builder().CreateFPToUI(f.v, llvm::Type::getInt32Ty(SSAScope::context()), SSAScope::hint());
+	else
+		v = SSAScope::builder().CreateFPToSI(f.v, llvm::Type::getInt32Ty(SSAScope::context()), SSAScope::hint());
 }
 
 SSAInt::SSAInt(llvm::Value *v)
@@ -63,6 +66,11 @@ SSAInt SSAInt::MAX(SSAInt a, SSAInt b)
 	return SSAInt::from_llvm(SSAScope::builder().CreateSelect((a > b).v, a.v, b.v, SSAScope::hint()));
 }
 
+SSAInt SSAInt::clamp(SSAInt a, SSAInt b, SSAInt c)
+{
+	return SSAInt::MAX(SSAInt::MIN(a, c), b);
+}
+
 SSAInt SSAInt::add(SSAInt b, bool no_unsigned_wrap, bool no_signed_wrap)
 {
 	return SSAInt::from_llvm(SSAScope::builder().CreateAdd(v, b.v, SSAScope::hint(), no_unsigned_wrap, no_signed_wrap));

src/r_compiler/ssa/ssa_int.h

@@ -32,13 +32,14 @@ class SSAInt
 public:
 	SSAInt();
 	explicit SSAInt(int constant);
-	SSAInt(SSAFloat f);
+	SSAInt(SSAFloat f, bool uint);
 	explicit SSAInt(llvm::Value *v);
 	static SSAInt from_llvm(llvm::Value *v) { return SSAInt(v); }
 	static llvm::Type *llvm_type();
 
 	static SSAInt MIN(SSAInt a, SSAInt b);
 	static SSAInt MAX(SSAInt a, SSAInt b);
+	static SSAInt clamp(SSAInt a, SSAInt b, SSAInt c);
 
 	SSAInt add(SSAInt b, bool no_unsigned_wrap, bool no_signed_wrap);
 	SSAInt ashr(int bits);

src/r_compiler/ssa/ssa_scope.cpp

@@ -56,7 +56,7 @@ llvm::IRBuilder<> &SSAScope::builder()
 
 llvm::Function *SSAScope::intrinsic(llvm::Intrinsic::ID id, llvm::ArrayRef<llvm::Type *> parameter_types)
 {
-	llvm::Function *func = module()->getFunction(llvm::Intrinsic::getName(id));
+	llvm::Function *func = module()->getFunction(llvm::Intrinsic::getName(id, parameter_types));
 	if (func == 0)
 		func = llvm::Function::Create(llvm::Intrinsic::getType(context(), id, parameter_types), llvm::Function::ExternalLinkage, llvm::Intrinsic::getName(id, parameter_types), module());
 	return func;

src/r_poly.cpp

@@ -222,13 +222,13 @@ void RenderPolyBsp::RenderPlane(subsector_t *sub, uint32_t subsectorDepth, bool
 	if (!isSky)
 	{
 		args.SetTexture(tex);
-		PolyTriangleDrawer::draw(args, PolyDrawVariant::Draw);
+		PolyTriangleDrawer::draw(args, TriDrawVariant::Draw);
-		PolyTriangleDrawer::draw(args, PolyDrawVariant::Stencil);
+		PolyTriangleDrawer::draw(args, TriDrawVariant::Stencil);
 	}
 	else
 	{
 		args.stencilwritevalue = 255;
-		PolyTriangleDrawer::draw(args, PolyDrawVariant::Stencil);
+		PolyTriangleDrawer::draw(args, TriDrawVariant::Stencil);
 
 		for (uint32_t i = 0; i < sub->numlines; i++)
 		{
@@ -313,7 +313,7 @@ void RenderPolyBsp::RenderPlane(subsector_t *sub, uint32_t subsectorDepth, bool
 
 			args.vinput = wallvert;
 			args.vcount = 4;
-			PolyTriangleDrawer::draw(args, PolyDrawVariant::Stencil);
+			PolyTriangleDrawer::draw(args, TriDrawVariant::Stencil);
 		}
 	}
 }
@@ -568,7 +568,7 @@ void RenderPolyBsp::AddSprite(AActor *thing, subsector_t *sub, uint32_t subsecto
 	args.stenciltestvalue = 0;
 	args.stencilwritevalue = 1;
 	args.SetTexture(tex);
-	PolyTriangleDrawer::draw(args, PolyDrawVariant::DrawSubsector);
+	PolyTriangleDrawer::draw(args, TriDrawVariant::DrawSubsector);
 }
 
 void RenderPolyBsp::AddWallSprite(AActor *thing, subsector_t *sub, uint32_t subsectorDepth)
@@ -1234,8 +1234,8 @@ void RenderPolyWall::Render(const TriMatrix &worldToClip)
 	args.stencilwritevalue = 1;
 	args.SetTexture(tex);
 
-	PolyTriangleDrawer::draw(args, PolyDrawVariant::Draw);
+	PolyTriangleDrawer::draw(args, TriDrawVariant::Draw);
-	PolyTriangleDrawer::draw(args, PolyDrawVariant::Stencil);
+	PolyTriangleDrawer::draw(args, TriDrawVariant::Stencil);
 }
 
 FTexture *RenderPolyWall::GetTexture()
@@ -1567,7 +1567,7 @@ void PolySkyDome::RenderRow(PolyDrawArgs &args, int row)
 	args.vcount = mPrimStart[row + 1] - mPrimStart[row];
 	args.mode = TriangleDrawMode::Strip;
 	args.ccw = false;
-	PolyTriangleDrawer::draw(args, PolyDrawVariant::Draw);
+	PolyTriangleDrawer::draw(args, TriDrawVariant::Draw);
 }
 
 void PolySkyDome::RenderCapColorRow(PolyDrawArgs &args, FTexture *skytex, int row, bool bottomCap)
@@ -1581,7 +1581,7 @@ void PolySkyDome::RenderCapColorRow(PolyDrawArgs &args, FTexture *skytex, int ro
 	args.mode = TriangleDrawMode::Fan;
 	args.ccw = bottomCap;
 	args.solidcolor = solid;
-	PolyTriangleDrawer::draw(args, PolyDrawVariant::Fill);
+	PolyTriangleDrawer::draw(args, TriDrawVariant::Fill);
 }
 
 void PolySkyDome::Render(const TriMatrix &worldToClip)

src/r_poly.h

@@ -106,7 +106,7 @@ public:
 	void Render();
 	void RenderScreenSprites();
 
-	static const uint32_t SkySubsectorDepth = 0xffffffff;
+	static const uint32_t SkySubsectorDepth = 0x7fffffff;
 
 private:
 	void RenderNode(void *node);

src/r_poly_triangle.cpp

@@ -40,7 +40,7 @@
 #include <immintrin.h>
 #endif
 
-void PolyTriangleDrawer::draw(const PolyDrawArgs &args, PolyDrawVariant variant)
+void PolyTriangleDrawer::draw(const PolyDrawArgs &args, TriDrawVariant variant)
 {
 	if (r_swtruecolor)
 		DrawerCommandQueue::QueueCommand<DrawPolyTrianglesCommand>(args, variant);
@@ -48,19 +48,20 @@ void PolyTriangleDrawer::draw(const PolyDrawArgs &args, PolyDrawVariant variant)
 		draw_arrays(args, variant, nullptr);
 }
 
-void PolyTriangleDrawer::draw_arrays(const PolyDrawArgs &drawargs, PolyDrawVariant variant, WorkerThreadData *thread)
+void PolyTriangleDrawer::draw_arrays(const PolyDrawArgs &drawargs, TriDrawVariant variant, WorkerThreadData *thread)
 {
 	if (drawargs.vcount < 3)
 		return;
 
+	auto llvm = LLVMDrawers::Instance();
 	void(*drawfunc)(const TriDrawTriangleArgs *, WorkerThreadData *);
 	switch (variant)
 	{
 	default:
-	case PolyDrawVariant::Draw: drawfunc = r_swtruecolor ? ScreenPolyTriangleDrawer::draw32 : ScreenPolyTriangleDrawer::draw; break;
+	case TriDrawVariant::Draw: drawfunc = r_swtruecolor ? ScreenPolyTriangleDrawer::draw32 : ScreenPolyTriangleDrawer::draw; break;
-	case PolyDrawVariant::Fill: drawfunc = r_swtruecolor ? ScreenPolyTriangleDrawer::fill32 : ScreenPolyTriangleDrawer::fill; break;
+	case TriDrawVariant::Fill: drawfunc = r_swtruecolor ? ScreenPolyTriangleDrawer::fill32 : ScreenPolyTriangleDrawer::fill; break;
-	case PolyDrawVariant::DrawSubsector: drawfunc = r_swtruecolor ? ScreenPolyTriangleDrawer::drawsubsector32 : ScreenPolyTriangleDrawer::draw; break;
+	case TriDrawVariant::DrawSubsector: drawfunc = r_swtruecolor ? llvm->TriDrawSubsector32 : llvm->TriDrawSubsector8; break;
-	case PolyDrawVariant::Stencil: drawfunc = ScreenPolyTriangleDrawer::stencil; break;
+	case TriDrawVariant::Stencil: drawfunc = ScreenPolyTriangleDrawer::stencil; break;
 	}
 
 	TriDrawTriangleArgs args;
@@ -1596,7 +1597,7 @@ float ScreenPolyTriangleDrawer::grady(float x0, float y0, float x1, float y1, fl
 
 /////////////////////////////////////////////////////////////////////////////
 
-DrawPolyTrianglesCommand::DrawPolyTrianglesCommand(const PolyDrawArgs &args, PolyDrawVariant variant)
+DrawPolyTrianglesCommand::DrawPolyTrianglesCommand(const PolyDrawArgs &args, TriDrawVariant variant)
 	: args(args), variant(variant)
 {
 }

src/r_poly_triangle.h

@@ -28,14 +28,6 @@
 
 struct TriDrawTriangleArgs;
 
-enum class PolyDrawVariant
-{
-	Draw,
-	Fill,
-	DrawSubsector,
-	Stencil,
-};
-
 class PolyDrawArgs
 {
 public:
@@ -69,11 +61,11 @@ public:
 class PolyTriangleDrawer
 {
 public:
-	static void draw(const PolyDrawArgs &args, PolyDrawVariant variant);
+	static void draw(const PolyDrawArgs &args, TriDrawVariant variant);
 
 private:
 	static TriVertex shade_vertex(const TriUniforms &uniforms, TriVertex v);
-	static void draw_arrays(const PolyDrawArgs &args, PolyDrawVariant variant, WorkerThreadData *thread);
+	static void draw_arrays(const PolyDrawArgs &args, TriDrawVariant variant, WorkerThreadData *thread);
 	static void draw_shaded_triangle(const TriVertex *vertices, bool ccw, TriDrawTriangleArgs *args, WorkerThreadData *thread, void(*drawfunc)(const TriDrawTriangleArgs *, WorkerThreadData *));
 	static bool cullhalfspace(float clipdistance1, float clipdistance2, float &t1, float &t2);
 	static void clipedge(const TriVertex *verts, TriVertex *clippedvert, int &numclipvert);
@@ -245,14 +237,14 @@ private:
 class DrawPolyTrianglesCommand : public DrawerCommand
 {
 public:
-	DrawPolyTrianglesCommand(const PolyDrawArgs &args, PolyDrawVariant variant);
+	DrawPolyTrianglesCommand(const PolyDrawArgs &args, TriDrawVariant variant);
 
 	void Execute(DrawerThread *thread) override;
 	FString DebugInfo() override;
 
 private:
 	PolyDrawArgs args;
-	PolyDrawVariant variant;
+	TriDrawVariant variant;
 };
 
 #endif