Fully implemented codegen for DrawSpan

src/r_compiler/fixedfunction/fixedfunction.h

@@ -1,6 +1,7 @@
 
 #pragma once
 
+#include "r_compiler/ssa/ssa_value.h"
 #include "r_compiler/ssa/ssa_vec4f.h"
 #include "r_compiler/ssa/ssa_vec4i.h"
 #include "r_compiler/ssa/ssa_vec8s.h"
@@ -84,16 +85,9 @@ public:
 	SSAInt desaturate;
 };
 
-class FixedFunction
+class DrawerCodegen
 {
 public:
-	FixedFunction();
-
-	void(*DrawSpan)(const RenderArgs *) = nullptr;
-
-private:
-	void CodegenDrawSpan();
-
 	// LightBgra
 	SSAInt calc_light_multiplier(SSAInt light);
 	SSAVec4i shade_pal_index_simple(SSAInt index, SSAInt light, SSAUBytePtr basecolors);
@@ -111,89 +105,57 @@ private:
 	// SampleBgra
 	SSAVec4i sample_linear(SSAUBytePtr col0, SSAUBytePtr col1, SSAInt texturefracx, SSAInt texturefracy, SSAInt one, SSAInt height);
 	SSAVec4i sample_linear(SSAUBytePtr texture, SSAInt xfrac, SSAInt yfrac, SSAInt xbits, SSAInt ybits);
+};
+
+class DrawSpanCodegen : public DrawerCodegen
+{
+public:
+	void Generate(SSAValue args);
+
+private:
+	void LoopShade(bool isSimpleShade);
+	void LoopFilter(bool isSimpleShade, bool isNearestFilter);
+	SSAInt Loop4x(bool isSimpleShade, bool isNearestFilter, bool is64x64);
+	void Loop(SSAInt start, bool isSimpleShade, bool isNearestFilter, bool is64x64);
+	SSAVec4i Sample(SSAInt xfrac, SSAInt yfrac, bool isNearestFilter, bool is64x64);
+
+	SSAStack<SSAInt> stack_index, stack_xfrac, stack_yfrac;
+
+	SSAUBytePtr destorg;
+	SSAUBytePtr source;
+	SSAInt destpitch;
+	SSAInt xstep;
+	SSAInt ystep;
+	SSAInt x1;
+	SSAInt x2;
+	SSAInt y;
+	SSAInt xbits;
+	SSAInt ybits;
+	SSAInt light;
+	SSAInt srcalpha;
+	SSAInt destalpha;
+	SSAInt count;
+	SSAUBytePtr data;
+	SSAInt yshift;
+	SSAInt xshift;
+	SSAInt xmask;
+	SSABool is_64x64;
+	SSABool is_simple_shade;
+	SSABool is_nearest_filter;
+	SSAShadeConstants shade_constants;
+};
+
+class FixedFunction
+{
+public:
+	FixedFunction();
+
+	void(*DrawSpan)(const RenderArgs *) = nullptr;
+
+private:
+	void CodegenDrawSpan();
 
 	static llvm::Type *GetRenderArgsStruct(llvm::LLVMContext &context);
 
 	RenderProgram mProgram;
 };
-
-#if 0
-
-class GlslProgram;
-class GlslCodeGen;
-
-class GlslFixedFunction
-{
-public:
-	GlslFixedFunction(GlslProgram &program, GlslCodeGen &vertex_codegen, GlslCodeGen &fragment_codegen);
-	void codegen();
-	static llvm::Type *get_sampler_struct(llvm::LLVMContext &context);
-
-private:
-	void codegen_draw_triangles(int num_vertex_in, int num_vertex_out);
-	void codegen_calc_window_positions();
-	void codegen_calc_polygon_face_direction();
-	void codegen_calc_polygon_y_range();
-	void codegen_update_polygon_edge();
-	void codegen_texture();
-	void codegen_normalize();
-	void codegen_reflect();
-	void codegen_max();
-	void codegen_pow();
-	void codegen_dot();
-	void codegen_mix();
-
-	struct OuterData
-	{
-		OuterData() : sampler() { }
-
-		SSAInt start;
-		SSAInt end;
-		SSAInt input_width;
-		SSAInt input_height;
-		SSAInt output_width;
-		SSAInt output_height;
-		SSAUBytePtr input_pixels;
-		SSAUBytePtr output_pixels_line;
-
-		SSAVec4fPtr sse_left_varying_in;
-		SSAVec4fPtr sse_right_varying_in;
-		int num_varyings;
-		SSAVec4f viewport_x;
-		SSAVec4f viewport_rcp_half_width;
-		SSAVec4f dx;
-		SSAVec4f dw;
-		SSAVec4f v1w;
-		SSAVec4f v1x;
-
-		llvm::Value *sampler;
-	};
-
-	void render_polygon(
-		SSAInt input_width,
-		SSAInt input_height,
-		SSAUBytePtr input_data,
-		SSAInt output_width,
-		SSAInt output_height,
-		SSAUBytePtr output_data,
-		SSAInt viewport_x,
-		SSAInt viewport_y,
-		SSAInt viewport_width,
-		SSAInt viewport_height,
-		SSAInt num_vertices,
-		std::vector<SSAVec4fPtr> fragment_ins,
-		SSAInt core,
-		SSAInt num_cores);
-
-	void codegen_render_scanline(int num_varyings);
-	void process_first_pixels(OuterData &outer_data, SSAStack<SSAInt> &stack_x, SSAStack<SSAVec4f> &stack_xnormalized);
-	void process_last_pixels(OuterData &outer_data, SSAStack<SSAInt> &stack_x, SSAStack<SSAVec4f> &stack_xnormalized);
-	void inner_block(OuterData &data, SSAVec4f xnormalized, SSAVec4f *out_frag_colors);
-	void blend(SSAVec4f frag_colors[4], SSAVec16ub &dest);
-
-	GlslProgram &program;
-	GlslCodeGen &vertex_codegen;
-	GlslCodeGen &fragment_codegen;
-};
-
-#endif

src/r_compiler/ssa/ssa_short.cpp

@@ -32,6 +32,11 @@ llvm::Type *SSAShort::llvm_type()
 	return llvm::Type::getInt16Ty(SSAScope::context());
 }
 
+SSAInt SSAShort::zext_int()
+{
+	return SSAInt::from_llvm(SSAScope::builder().CreateZExt(v, SSAInt::llvm_type(), SSAScope::hint()));
+}
+
 SSAShort operator+(const SSAShort &a, const SSAShort &b)
 {
 	return SSAShort::from_llvm(SSAScope::builder().CreateAdd(a.v, b.v, SSAScope::hint()));

src/r_compiler/ssa/ssa_short.h

@@ -17,6 +17,8 @@ public:
 	static SSAShort from_llvm(llvm::Value *v) { return SSAShort(v); }
 	static llvm::Type *llvm_type();
 
+	SSAInt zext_int();
+
 	llvm::Value *v;
 };
 

src/r_compiler/ssa/ssa_vec4i.cpp

@@ -49,6 +49,18 @@ SSAVec4i::SSAVec4i(SSAInt i)
 	v = SSAScope::builder().CreateShuffleVector(SSAScope::builder().CreateBitCast(i.v, m1xi32type, SSAScope::hint()), llvm::UndefValue::get(m1xi32type), mask, SSAScope::hint());
 }
 
+SSAVec4i::SSAVec4i(SSAInt i0, SSAInt i1, SSAInt i2, SSAInt i3)
+: v(0)
+{
+	std::vector<llvm::Constant*> constants;
+	constants.resize(4, llvm::ConstantInt::get(SSAScope::context(), llvm::APInt(32, 0, true)));
+	v = llvm::ConstantVector::get(constants);
+	v = SSAScope::builder().CreateInsertElement(v, i0.v, (uint64_t)0, SSAScope::hint());
+	v = SSAScope::builder().CreateInsertElement(v, i1.v, (uint64_t)1, SSAScope::hint());
+	v = SSAScope::builder().CreateInsertElement(v, i2.v, (uint64_t)2, SSAScope::hint());
+	v = SSAScope::builder().CreateInsertElement(v, i3.v, (uint64_t)3, SSAScope::hint());
+}
+
 SSAVec4i::SSAVec4i(SSAVec4f f32)
 : v(0)
 {

src/r_compiler/ssa/ssa_vec4i.h

@@ -16,6 +16,7 @@ public:
 	SSAVec4i(int constant);
 	SSAVec4i(int constant0, int constant1, int constant2, int constant3);
 	SSAVec4i(SSAInt i);
+	SSAVec4i(SSAInt i0, SSAInt i1, SSAInt i2, SSAInt i3);
 	explicit SSAVec4i(llvm::Value *v);
 	SSAVec4i(SSAVec4f f32);
 	SSAInt operator[](SSAInt index);

src/r_draw_rgba.cpp

@@ -300,50 +300,43 @@ void DrawerCommandQueue::StopThreads()
 
 /////////////////////////////////////////////////////////////////////////////
 
-class DrawSpanFFCommand : public DrawerCommand
+class DrawSpanLLVMCommand : public DrawerCommand
 {
-	fixed_t _xfrac;
+	RenderArgs args;
-	fixed_t _yfrac;
-	fixed_t _xstep;
-	fixed_t _ystep;
-	int _x1;
-	int _x2;
-	int _y;
-	int _xbits;
-	int _ybits;
-	BYTE * RESTRICT _destorg;
-
-	const uint32_t * RESTRICT _source;
-	uint32_t _light;
-	ShadeConstants _shade_constants;
-	bool _nearest_filter;
-
-	uint32_t _srcalpha;
-	uint32_t _destalpha;
-
 	FixedFunction *_ff;
 
 public:
-	DrawSpanFFCommand()
+	DrawSpanLLVMCommand()
 	{
-		_xfrac = ds_xfrac;
+		args.xfrac = ds_xfrac;
-		_yfrac = ds_yfrac;
+		args.yfrac = ds_yfrac;
-		_xstep = ds_xstep;
+		args.xstep = ds_xstep;
-		_ystep = ds_ystep;
+		args.ystep = ds_ystep;
-		_x1 = ds_x1;
+		args.x1 = ds_x1;
-		_x2 = ds_x2;
+		args.x2 = ds_x2;
-		_y = ds_y;
+		args.y = ds_y;
-		_xbits = ds_xbits;
+		args.xbits = ds_xbits;
-		_ybits = ds_ybits;
+		args.ybits = ds_ybits;
-		_destorg = dc_destorg;
+		args.destorg = (uint32_t*)dc_destorg;
-
+		args.destpitch = dc_pitch;
-		_source = (const uint32_t*)ds_source;
+		args.source = (const uint32_t*)ds_source;
-		_light = LightBgra::calc_light_multiplier(ds_light);
+		args.light = LightBgra::calc_light_multiplier(ds_light);
-		_shade_constants = ds_shade_constants;
+		args.light_red = ds_shade_constants.light_red;
-		_nearest_filter = !SampleBgra::span_sampler_setup(_source, _xbits, _ybits, _xstep, _ystep, ds_source_mipmapped);
+		args.light_green = ds_shade_constants.light_green;
-
+		args.light_blue = ds_shade_constants.light_blue;
-		_srcalpha = dc_srcalpha >> (FRACBITS - 8);
+		args.light_alpha = ds_shade_constants.light_alpha;
-		_destalpha = dc_destalpha >> (FRACBITS - 8);
+		args.fade_red = ds_shade_constants.fade_red;
+		args.fade_green = ds_shade_constants.fade_green;
+		args.fade_blue = ds_shade_constants.fade_blue;
+		args.fade_alpha = ds_shade_constants.fade_alpha;
+		args.desaturate = ds_shade_constants.desaturate;
+		args.srcalpha = dc_srcalpha >> (FRACBITS - 8);
+		args.destalpha = dc_destalpha >> (FRACBITS - 8);
+		args.flags = 0;
+		if (ds_shade_constants.simple_shade)
+			args.flags |= RenderArgs::simple_shade;
+		if (!SampleBgra::span_sampler_setup(args.source, args.xbits, args.ybits, args.xstep, args.ystep, ds_source_mipmapped))
+			args.flags |= RenderArgs::nearest_filter;
 
 		static FixedFunction ff;
 		_ff = &ff;
@@ -351,25 +344,8 @@ public:
 
 	void Execute(DrawerThread *thread) override
 	{
-		if (thread->skipped_by_thread(_y))
+		if (thread->skipped_by_thread(args.y))
 			return;
-
-		RenderArgs args;
-		args.destorg = (uint32_t *)_destorg;
-		args.source = _source;
-		args.destpitch = dc_pitch;
-		args.xfrac = _xfrac;
-		args.yfrac = _yfrac;
-		args.xstep = _xstep;
-		args.ystep = _ystep;
-		args.x1 = _x1;
-		args.x2 = _x2;
-		args.y = _y;
-		args.xbits = _xbits;
-		args.ybits = _ybits;
-		args.light = _light;
-		args.srcalpha = _srcalpha;
-		args.destalpha = _destalpha;
 		_ff->DrawSpan(&args);
 	}
 };
@@ -2777,7 +2753,7 @@ void R_DrawRevSubClampTranslatedColumn_rgba()
 
 void R_DrawSpan_rgba()
 {
-	DrawerCommandQueue::QueueCommand<DrawSpanFFCommand>();
+	DrawerCommandQueue::QueueCommand<DrawSpanLLVMCommand>();
 /*
 #ifdef NO_SSE
 	DrawerCommandQueue::QueueCommand<DrawSpanRGBACommand>();