#include "externals.hh"
#include "control.hh"
#include "globals.hh"
#include "sync.hh"
#include "profiling.hh"

using namespace ops;
using namespace ebcl;


/*= OPCODE INFORMATIONS ========================================================*/

namespace {

struct T_OpInfo
{
	char const* name;
	uint32_t nArgs;
	int32_t sdMain{ 0 };
	int32_t sdFPU{ 0 };
	uint32_t fpuReq{ 0 };

	template< typename ... Mods >
	constexpr T_OpInfo( char const* name ,
			uint32_t nArgs ,
			Mods&&... mods );

	constexpr T_OpInfo( char const* name );
};


template< typename ... Mods >
constexpr void ApplyMods_( T_OpInfo& op , Mods&&... mods );

template< >
constexpr void ApplyMods_( T_OpInfo& )
{ }

template< typename M0 , typename ... Mods >
constexpr void ApplyMods_(
		T_OpInfo& op ,
		M0&& mod ,
		Mods&&... mods )
{
	mod.applyTo( op );
	ApplyMods_( op , std::forward< Mods >( mods ) ... );
}

template< typename ... Mods >
constexpr T_OpInfo::T_OpInfo(
		char const* name ,
		uint32_t nArgs ,
		Mods&&... mods )
	: name( name ) , nArgs( nArgs )
{
	ApplyMods_( *this , std::forward< Mods >( mods ) ... );
}

constexpr T_OpInfo::T_OpInfo( char const* name )
	: T_OpInfo( name , 0 )
{}


struct OpStackMain
{
	int32_t value;
	explicit constexpr OpStackMain( int32_t v ) : value( v ) { }
	constexpr void applyTo( T_OpInfo& op )
		{ op.sdMain = value; }
};

struct OpStackFPU
{
	int32_t value;
	explicit constexpr OpStackFPU( int32_t v ) : value( v ) { }
	constexpr void applyTo( T_OpInfo& op )
		{ op.sdFPU = value; }
};


static T_KeyValueTable< E_OpType , T_OpInfo > OpInfoTable_{ ([]() {
	T_KeyValueTable< E_OpType , T_OpInfo > infos;

	infos.add( E_OpType::OP_END , T_OpInfo{ "end" } );
	//
	infos.add( E_OpType::OP_CALL , T_OpInfo{ "call" , 1 } );
	infos.add( E_OpType::OP_RET , T_OpInfo{ "ret" , 1 , OpStackMain{ -1 } } );
	//
	infos.add( E_OpType::OP_SKIP , T_OpInfo{ "skip" , 1 } );
	infos.add( E_OpType::OP_COND_SKIP , T_OpInfo{ "cond-skip" , 2 } );
	//
	infos.add( E_OpType::OP_RES_STACK , T_OpInfo{ "res-stack" , 1 } );
	infos.add( E_OpType::OP_PUSH , T_OpInfo{ "push" , 0 , OpStackMain{ 1 } } );
	infos.add( E_OpType::OP_POP , T_OpInfo{ "pop" , 1 } );
	infos.add( E_OpType::OP_DUP , T_OpInfo{ "dup" , 1 , OpStackMain{ 1 } } );
	//
	infos.add( E_OpType::OP_LOAD , T_OpInfo{ "load" , 1 } );
	infos.add( E_OpType::OP_SLOAD , T_OpInfo{ "load-stack" , 1 } );
	infos.add( E_OpType::OP_CONST , T_OpInfo{ "const" , 1 } );
	//
	infos.add( E_OpType::OP_FP_LOAD , T_OpInfo{ "fp-load" , 1 , OpStackFPU{ 1 } } );
	infos.add( E_OpType::OP_FP_SLOAD , T_OpInfo{ "fp-load-stack" , 1 , OpStackFPU{ 1 } } );
	infos.add( E_OpType::OP_FP_STORE , T_OpInfo{ "fp-store" , 1 , OpStackFPU{ -1 } } );
	infos.add( E_OpType::OP_FP_SSTORE , T_OpInfo{ "fp-store-stack" , 1 , OpStackFPU{ -1 } } );
	infos.add( E_OpType::OP_FP_SSTORE_INT , T_OpInfo{ "fp-store-stack-int" , 1 , OpStackFPU{ -1 } } );
	//
	infos.add( E_OpType::OP_FP_CMP , T_OpInfo{ "fp-cmp" , 1 , OpStackFPU{ -1 } } );
	infos.add( E_OpType::OP_FP_ADD , T_OpInfo{ "fp-add" , 0 , OpStackFPU{ -1 } } );
	infos.add( E_OpType::OP_FP_SUB , T_OpInfo{ "fp-sub" , 0 , OpStackFPU{ -1 } } );
	infos.add( E_OpType::OP_FP_MUL , T_OpInfo{ "fp-mul" , 0 , OpStackFPU{ -1 } } );
	infos.add( E_OpType::OP_FP_DIV , T_OpInfo{ "fp-div" , 0 , OpStackFPU{ -1 } } );
	infos.add( E_OpType::OP_FP_POW , T_OpInfo{ "fp-pow" , 0 , OpStackFPU{ -1 } } );
	//
	infos.add( E_OpType::OP_FP_NEG , T_OpInfo{ "fp-neg" , 0 } );
	infos.add( E_OpType::OP_FP_INV , T_OpInfo{ "fp-inv" , 0 } );
	infos.add( E_OpType::OP_FP_NOT , T_OpInfo{ "fp-not" , 0 } );
	infos.add( E_OpType::OP_FP_SIN , T_OpInfo{ "fp-sin" , 0 } );
	infos.add( E_OpType::OP_FP_COS , T_OpInfo{ "fp-cos" , 0 } );
	infos.add( E_OpType::OP_FP_TAN , T_OpInfo{ "fp-tan" , 0 } );
	infos.add( E_OpType::OP_FP_SQRT , T_OpInfo{ "fp-sqrt" , 0 } );
	infos.add( E_OpType::OP_FP_EXP , T_OpInfo{ "fp-exp" , 0 } );
	infos.add( E_OpType::OP_FP_LN , T_OpInfo{ "fp-ln" , 0 } );
	//
	infos.add( E_OpType::OP_INIT_PIPELINE , T_OpInfo{ "pipeline" , 1 , OpStackMain{ -1 } } );
	infos.add( E_OpType::OP_INIT_PROGRAM , T_OpInfo{ "program" , 1 , OpStackMain{ -1 } } );
	infos.add( E_OpType::OP_INIT_TEXTURE , T_OpInfo{ "texture" , 2 , OpStackMain{ -3 } } );
	infos.add( E_OpType::OP_INIT_SAMPLER , T_OpInfo{ "sampler" , 2 , OpStackMain{ -3 } } );
	infos.add( E_OpType::OP_FB_ATTACH , T_OpInfo{ "fb-attach" , 2 , OpStackMain{ -1 } } );
	//
	infos.add( E_OpType::OP_USE_FRAMEBUFFER , T_OpInfo{ "use-framebuffer" , 0 , OpStackMain{ -1 } } );
	infos.add( E_OpType::OP_FB_TOGGLE , T_OpInfo{ "fb-toggle" , 2 } );
	infos.add( E_OpType::OP_USE_PIPELINE , T_OpInfo{ "use-pipeline" , 0 , OpStackMain{ -1 } } );
	infos.add( E_OpType::OP_USE_PROGRAM , T_OpInfo{ "use-program" , 0 , OpStackMain{ -1 } } );
	infos.add( E_OpType::OP_USE_TEXTURE , T_OpInfo{ "use-texture" , 1 , OpStackMain{ -2 } } );
	infos.add( E_OpType::OP_UNIFORMS , T_OpInfo{ "uniforms" , 2 , OpStackMain{ -2 } } );
	infos.add( E_OpType::OP_VIEWPORT , T_OpInfo{ "viewport" , 0 , OpStackMain{ -4 } } );
	//
	infos.add( E_OpType::OP_FULLSCREEN , T_OpInfo{ "fullscreen" } );
	infos.add( E_OpType::OP_CLEAR , T_OpInfo{ "clear" , 0 , OpStackMain{ -4 } } );
	//
	infos.add( E_OpType::OP_UI_PENTER , T_OpInfo{ "ui-prof-enter" , 1 } );
	infos.add( E_OpType::OP_UI_PEXIT , T_OpInfo{ "ui-prof-exit" } );
	infos.add( E_OpType::OP_UI_INPUT_DFT , T_OpInfo{ "ui-input-default" , 2 } );
	infos.add( E_OpType::OP_UI_ODBG , T_OpInfo{ "ui-odbg" , 2 , OpStackMain{ -1 } } );

	return infos;
})( ) };

} // namespace

uint32_t ops::ArgumentsFor(
		const E_OpType op ) noexcept
{
	assert( OpInfoTable_.contains( op ) );
	return OpInfoTable_.get( op )->nArgs;
}

int32_t ops::DeltaMainStack(
		const E_OpType op ) noexcept
{
	assert( OpInfoTable_.contains( op ) );
	return OpInfoTable_.get( op )->sdMain;
}

int32_t ops::DeltaFPUStack(
		const E_OpType op ) noexcept
{
	assert( OpInfoTable_.contains( op ) );
	return OpInfoTable_.get( op )->sdFPU;
}


/*= STRING FORMATTING ==========================================================*/

T_StringBuilder& ops::operator<<(
		T_StringBuilder& sb ,
		const E_OpType et )
{
	assert( OpInfoTable_.contains( et ) );
	sb << OpInfoTable_.get( et )->name;
	return sb;
}

T_StringBuilder& ops::operator<<(
		T_StringBuilder& sb ,
		T_Op const& op )
{
	sb << op.op;
	const auto args{ OpInfoTable_.get( op.op )->nArgs };
	for ( auto i = 0u ; i < args ; i ++ ) {
		sb << ' ' << op.args[ i ];
	}
	return sb;
}


/*= EMULATOR ===================================================================*/

X_OpFailure::X_OpFailure(
		T_Op const& op ,
		T_String error ) noexcept
	: op_( &op ) , error_( std::move( error ) )
{
	T_StringBuilder sb;
	sb << "operation (" << op << ") failed; source: " << op.location
		<< '\0';
	fullMessage_ = std::move( sb );
}

char const* X_OpFailure::what( ) const noexcept
{
	return fullMessage_.data( );
}

/*------------------------------------------------------------------------------*/

T_OpContext::T_OpContext(
		T_OpProgram& program ) noexcept
	: program( program )
{
	stack.ensureCapacity( stack.growth( ) );

	const auto nc{ program.constants.size( ) };
	const auto ts{ 3 + nc + program.nVariables
		+ program.nPrograms + program.nPipelines
		+ program.nSamplers + program.nTextures };
	values.resize( ts );
	initialInputs.resize( program.inputs.size( ) );

	for ( auto i = 0u ; i < nc ; i ++ ) {
		values[ i + 3 ] = program.constants[ nc ];
	}
}

void T_OpContext::run(
		const E_RunTarget target ,
		const float time ,
		const float width ,
		const float height )
{
	assert( !aborted );
	x87sp = 0;
	instrPtr = program.ops.firstOf( target == R_INIT ? program.init : program.frame );
	values[ 0 ] = time;
	values[ 1 ] = width;
	values[ 2 ] = height;
	stack.clear( );
	stack.add( 0xffffffff );

	while ( 0 /* FIXME */ ) {
//		if ( instrPtr >= program.ops.
		auto const& instr{ program.ops[ instrPtr ] };

		switch ( instr.op ) {

		    case OP_END:
			throw X_OpFailure{ instr , "invalid instruction" };

		    // --------------------------------------------------------------------------------

		    case OP_CALL:
			if ( program.ops.size( ) <= instr.args[ 0 ] ) {
				throw X_OpFailure{ instr , "invalid function index" };
			}
			stack.add( instrPtr + 1 );
			instrPtr = program.ops.firstOf( instr.args[ 0 ] );
			continue;

		    case OP_RET:
			ensureStack( instr , 1 + instr.args[ 0 ] );
			instrPtr = stack.last( ).u;
			stack.resize( stack.size( ) - instr.args[ 0 ] - 1 );
			continue;

		    case OP_SKIP:
			// FIXME: make sure we don't go past the end
			instrPtr += instr.args[ 0 ];
			break;

		    case OP_COND_SKIP:
			ensureStack( instr , 1 );
			if ( stack.last( ).u != instr.args[ 1 ] ) {
				// FIXME: make sure we don't go past the end
				instrPtr += instr.args[ 0 ];
			}
			break;

		    // --------------------------------------------------------------------------------

		    case OP_RES_STACK:
			stack.resize( stack.size( ) + instr.args[ 0 ] );
			break;

		    case OP_PUSH:
			stack.add( wreg );
			break;

		    case OP_POP:
			ensureStack( instr , instr.args[ 0 ] + 1 );
			stack.resize( stack.size( ) - instr.args[ 0 ] - 1 );
			break;

		    case OP_DUP:
			ensureStack( instr , instr.args[ 0 ] + 1 );
			stack.add( stack[ stack.size( ) - instr.args[ 0 ] - 1 ] );
			break;

		    // --------------------------------------------------------------------------------

		    case OP_LOAD:
			checkAddress( instr , instr.args[ 0 ] );
			wreg = values[ instr.args[ 0 ] ];
			break;

		    case OP_SLOAD:
			ensureStack( instr , instr.args[ 0 ] + 1 );
			wreg = stack[ stack.size( ) - instr.args[ 0 ] - 1 ];
			break;

		    case OP_CONST:
			wreg = instr.args[ 0 ];
			break;

		    // --------------------------------------------------------------------------------

		    case OP_FP_LOAD:
			ensureFpuStack( instr , 0 , 1 );
			checkAddress( instr , instr.args[ 0 ] );
			x87stack[ x87sp ++ ] = values[ instr.args[ 0 ] ].f;
			break;

		    case OP_FP_STORE:
			ensureFpuStack( instr , 1 , 0 );
			checkAddress( instr , instr.args[ 0 ] );
			values[ instr.args[ 0 ] ].f = x87stack[ -- x87sp ];
			break;

		    case OP_FP_SLOAD:
			ensureFpuStack( instr , 0 , 1 );
			ensureStack( instr , instr.args[ 0 ] + 1 );
			x87stack[ x87sp ++ ] = stack[ stack.size( ) - instr.args[ 0 ] - 1 ].f;
			break;

		    case OP_FP_SSTORE:
			ensureFpuStack( instr , 1 , 0 );
			ensureStack( instr , instr.args[ 0 ] + 1 );
			stack[ stack.size( ) - instr.args[ 0 ] - 1 ].f = x87stack[ -- x87sp ];
			break;

		    case OP_FP_SSTORE_INT:
			ensureFpuStack( instr , 1 , 0 );
			ensureStack( instr , instr.args[ 0 ] + 1 );
			stack[ stack.size( ) - instr.args[ 0 ] - 1 ].i = int32_t( x87stack[ -- x87sp ] );
			break;

		    // --------------------------------------------------------------------------------

		    case OP_FP_CMP:
		    {
			ensureFpuStack( instr , 2 , 0 );
			const auto v2( x87stack[ x87sp - 1 ] ) ,
					v1( x87stack[ x87sp - 2 ] );
			x87sp --;

			const bool rv{ ([&]( ) -> bool {
				switch ( instr.args[ 0 ] ) {
				    case 0: return v1 == v2;
				    case 1: return v1 != v2;
				    case 2: return v1 > v2;
				    case 3: return v1 >= v2;
				    case 4: return v1 < v2;
				    case 5: return v1 <= v2;
				}
				throw X_OpFailure( instr , "invalid operation" );
			})() };
			x87stack[ x87sp - 1 ] = rv ? 1 : 0;
			break;
		    }

		    case OP_FP_ADD:
		    {
			ensureFpuStack( instr , 2 , 0 );
			x87stack[ x87sp - 2 ] += x87stack[ x87sp - 1 ];
			x87sp --;
			break;
		    }

		    case OP_FP_SUB:
		    {
			ensureFpuStack( instr , 2 , 0 );
			x87stack[ x87sp - 2 ] -= x87stack[ x87sp - 1 ];
			x87sp --;
			break;
		    }

		    case OP_FP_MUL:
		    {
			ensureFpuStack( instr , 2 , 0 );
			x87stack[ x87sp - 2 ] *= x87stack[ x87sp - 1 ];
			x87sp --;
			break;
		    }

		    case OP_FP_DIV:
		    {
			ensureFpuStack( instr , 2 , 0 );
			if ( x87stack[ x87sp - 1 ] == 0 ) {
				throw X_OpFailure{ instr , "arithmetic error" };
			}
			x87stack[ x87sp - 2 ] *= x87stack[ x87sp - 1 ];
			x87sp --;
			break;
		    }

		    case OP_FP_POW:
		    {
			ensureFpuStack( instr , 2 , 0 );
			x87stack[ x87sp - 2 ] = pow( x87stack[ x87sp - 2 ] ,
					x87stack[ x87sp - 1 ] );
			x87sp --;
			break;
		    }

		    // --------------------------------------------------------------------------------

		    case OP_FP_NEG:
		    {
			ensureFpuStack( instr , 1 , 1 );
			x87stack[ x87sp - 1 ] = -x87stack[ x87sp - 1 ];
			break;
		    }

		    case OP_FP_INV:
		    {
			ensureFpuStack( instr , 1 , 1 );
			if ( x87stack[ x87sp - 1 ] == 0 ) {
				throw X_OpFailure{ instr , "arithmetic error" };
			}
			x87stack[ x87sp - 1 ] = 1.0 / x87stack[ x87sp - 1 ];
			break;
		    }

		    case OP_FP_NOT:
		    {
			ensureFpuStack( instr , 1 , 0 );
			x87stack[ x87sp - 1 ] = x87stack[ x87sp - 1 ] == 0 ? 1 : 0;
			break;
		    }

		    case OP_FP_SIN:
		    {
			ensureFpuStack( instr , 1 , 0 );
			x87stack[ x87sp - 1 ] = sin( x87stack[ x87sp - 1 ] );
			break;
		    }

		    case OP_FP_COS:
		    {
			ensureFpuStack( instr , 1 , 0 );
			x87stack[ x87sp - 1 ] = cos( x87stack[ x87sp - 1 ] );
			break;
		    }

		    case OP_FP_TAN:
		    {
			ensureFpuStack( instr , 1 , 1 );
			const auto c( cos( x87stack[ x87sp - 1 ] ) );
			if ( c == 0 ) {
				throw X_OpFailure{ instr , "arithmetic error" };
			}
			x87stack[ x87sp - 1 ] = sin( x87stack[ x87sp - 1 ] ) / c;
			break;
		    }

		    case OP_FP_SQRT:
		    {
			ensureFpuStack( instr , 1 , 0 );
			const auto v( x87stack[ x87sp - 1 ] );
			if ( v < 0 ) {
				throw X_OpFailure{ instr , "arithmetic error" };
			}
			x87stack[ x87sp - 1 ] = sqrt( v );
			break;
		    }

		    case OP_FP_EXP:
		    {
			ensureFpuStack( instr , 1 , 0 ); // FIXME: need FPU stack space
			x87stack[ x87sp - 1 ] = exp( x87stack[ x87sp - 1 ] );
			break;
		    }

		    case OP_FP_LN:
		    {
			ensureFpuStack( instr , 1 , 0 ); // FIXME: need FPU stack space
			x87stack[ x87sp - 1 ] = log( x87stack[ x87sp - 1 ] );
			break;
		    }

		    // --------------------------------------------------------------------------------

		    // TODO resource init

		    // --------------------------------------------------------------------------------

		    // TODO resource usage

		    case OP_UNIFORMS:
		    {
			ensureStack( instr , 3 + instr.args[ 0 ] );

			const auto ss( stack.size( ) );
			T_OpValue values[ 4 ];
			for ( auto i = 0u ; i <= instr.args[ 0 ] ; i ++ ) {
				values[ i ] = stack[ ss - 2 - i ];
			}
			// FIXME can't actually finish this, I'm stupid.

			break;
		    }

		    case OP_VIEWPORT:
		    {
			ensureStack( instr , 4 );
			const auto ss( stack.size( ) );
			glViewport( stack[ ss - 1 ].f , stack[ ss - 2 ].f ,
					stack[ ss - 3 ].f , stack[ ss - 4 ].f );
			stack.resize( ss - 4 );
			break;
		    }

		    // --------------------------------------------------------------------------------

		    case OP_FULLSCREEN:
			glDrawArrays( GL_TRIANGLE_STRIP , 0 , 4 );
			break;

		    case OP_CLEAR:
		    {
			ensureStack( instr , 4 );
			const auto ss( stack.size( ) );
			glClearColor( stack[ ss - 1 ].f , stack[ ss - 2 ].f ,
					stack[ ss - 3 ].f , stack[ ss - 4 ].f );
			glClear( GL_COLOR_BUFFER_BIT | GL_DEPTH_BUFFER_BIT );
			stack.resize( ss - 4 );
			break;
		    }

		    // --------------------------------------------------------------------------------

		    case OP_UI_PENTER:
		    {
			T_String const& section( program.uiStrings[ instr.args[ 0 ] ] );
			Globals::Profiler( ).start( section );
			profiling.add( section );
			break;
		    }

		    case OP_UI_PEXIT:
			Globals::Profiler( ).end( profiling.last( ) );
			profiling.removeLast( );
			break;

		    case OP_UI_INPUT_DFT:
			initialInputs[ instr.args[ 0 ] ] = values[ instr.args[ 1 ] ].f;
			break;

		}

		instrPtr ++;
	}
}

void T_OpContext::ensureStack(
		T_Op const& op ,
		const uint32_t min )
{
	if ( stack.size( ) < min ) {
		throw X_OpFailure{ op , "stack underrun" };
	}
}

void T_OpContext::ensureFpuStack(
		T_Op const& op ,
		const uint32_t minStacked ,
		const uint32_t minFree )
{
	if ( x87sp < minStacked ) {
		throw X_OpFailure{ op , "FPU stack underrun" };
	}
	if ( 8 - x87sp < minFree ) {
		throw X_OpFailure{ op , "FPU stack overflow" };
	}
}

void T_OpContext::checkAddress(
		T_Op const& op ,
		const uint32_t address )
{
	if ( address >= values.size( ) ) {
		throw X_OpFailure( op , "invalid access" );
	}
}