Optimizer - UD chains - WORKING (fo'real, man)

Horrible, spaghettish, slow, memory hungry but WORKING implementation of
use/define chains!!!!

c-opopt.cc

@@ -761,6 +761,265 @@ void BUDCLink_(
 	} , LL2 );
 }
 
+/*----------------------------------------------------------------------------*/
+
+struct T_BUDCDefSet_
+{
+	T_Set< uint32_t > set{ UseTag< ArrayBacked< 4 > >( ) };
+	T_AutoArray< uint8_t , 1 > bitmap;
+
+	explicit T_BUDCDefSet_( const uint32_t max ) noexcept
+	{
+		const auto nBytesRaw{ max >> 3 };
+		const auto nBytes{ nBytesRaw + ( ( max & 7 ) != 0 ? 1 : 0 ) };
+		for ( auto i = 0u ; i < nBytes ; i ++ ) {
+			bitmap.add( 0 );
+		}
+	}
+
+	void clear( ) noexcept
+	{
+		set.clear( );
+		for ( auto i = 0u ; i < bitmap.size( ) ; i ++ ) {
+			bitmap[ i ] = 0;
+		}
+	}
+
+	void add( const uint32_t item ) noexcept
+	{
+		const auto byte{ item >> 3 };
+		const auto bit{ item & 7 };
+		const auto mask{ 1 << bit };
+		if ( ( bitmap[ byte ] & mask ) == 0 ) {
+			bitmap[ byte ] |= mask;
+			set.add( item );
+		}
+	}
+
+	bool operator ==( T_BUDCDefSet_ const& other ) noexcept
+	{
+		if ( bitmap.size( ) != other.bitmap.size( ) ) {
+			return false;
+		}
+		for ( auto i = 0u ; i < bitmap.size( ) ; i ++ ) {
+			if ( bitmap[ i ] != other.bitmap[ i ] ) {
+				return false;
+			}
+		}
+		return true;
+	}
+};
+
+void BUDCWalkGraph_(
+		T_OptData&			data ,
+		T_OptData::T_VarUseDefine&	var ,
+		T_UDEPerInstr_&			udPerInstr ,
+		T_Optional< uint32_t >		fnIndex
+		) noexcept
+{
+	T_Set< uint32_t > changed{ UseTag< ArrayBacked< 16 > >( ) }; // FIXME alloc
+	T_Array< T_BUDCDefSet_ > defines; // FIXME alloc
+	defines.resize( data.ctrlFlowGraph.size( ) , T_BUDCDefSet_{ 0 } );
+
+	data.logger( [&]() {
+		T_StringBuilder sb;
+		sb << "Walking graph for " << var.var.name << " (";
+		switch ( var.var.type ) {
+		    case T_OptData::E_UDVarType::GLOBAL:
+			sb << "global";
+			break;
+		    case T_OptData::E_UDVarType::ARGUMENT:
+			sb << "argument of " << var.var.owner;
+			break;
+		    case T_OptData::E_UDVarType::LOCAL:
+			sb << "local variable of " << var.var.owner;
+			break;
+		}
+		sb << ')';
+		return sb;
+	} , LL1 );
+
+	auto const* const fn{ fnIndex
+		? &data.cfgFunctions[ *fnIndex ]
+		: nullptr };
+	auto const nFirst{ fn ? fn->first : 0u };
+	auto const nEnd{ fn ? ( fn->first + fn->count )
+		: data.ctrlFlowGraph.size( ) };
+#if 0
+	data.logger( [=](){
+		T_StringBuilder sb{ "Nodes " };
+		sb << nFirst << " ... " << ( nEnd - 1 );;
+		return sb;
+	} , LL2 );
+#endif
+	changed.add( nFirst );
+	for ( auto i = nEnd ; i > nFirst + 1 ; i -- ) {
+		changed.add( i - 1 );
+	}
+
+	while ( changed.size( ) ) {
+		const auto node{ changed[ 0 ] };
+#if 0
+		data.logger( [=](){
+			T_StringBuilder sb{ "Checking node " };
+			sb << node;
+			return sb;
+		} , LL2 );
+#endif
+		assert( node >= nFirst && node < nEnd );
+		auto const& cn{ *data.ctrlFlowGraph[ node ] };
+		changed.remove( node );
+
+		// Assemble the set of possible definitions at the start of
+		// the block.
+		T_BUDCDefSet_ sInput{ var.defines.size( ) };
+		const auto nIn{ cn.inbound.size( ) };
+		for ( auto i = 0u ; i < nIn ; i ++ ) {
+			auto const& ib{ cn.inbound[ i ] };
+			// Make sure the inbound link type is right (FLOW always
+			// is, BYPASS works in functions, CALL/RET at the global
+			// level)
+			if ( ib.type != ib.FLOW && (
+					( fnIndex && ib.type != ib.BYPASS )
+					|| ( !fnIndex && ib.type == ib.BYPASS ) ) ) {
+				continue;
+			}
+
+			// Get defines from node
+			auto const& sNode{ defines[ ib.target ] };
+			const auto nIbIn{ sNode.set.size( ) };
+			for ( auto j = 0u ; j < nIbIn ; j ++ ) {
+				sInput.add( sNode.set[ j ] );
+			}
+		}
+
+		if ( cn.instructions ) {
+			// Check for defines in the instructions
+			const auto is{ cn.instructions->first };
+			const auto ie{ is + cn.instructions->count };
+			for ( auto ii = is ; ii < ie ; ii ++ ) {
+				auto const* const irec{ udPerInstr.get( ii ) };
+				if ( !irec ) {
+					continue;
+				}
+				const auto nrec{ irec->size( ) };
+
+				for ( auto j = 0u ; j < nrec ; j ++ ) {
+					auto const& rec{ (*irec)[ j ] };
+					if ( rec.isUse || data.varUDChains[ rec.entry ].var != var.var ) {
+						continue;
+					}
+					sInput.clear( );
+					sInput.add( rec.index );
+				}
+			}
+		}
+
+		auto& sOut{ defines[ node ] };
+		if ( sOut == sInput ) {
+			continue;
+		}
+#if 0
+		data.logger( [&](){
+			T_StringBuilder sb;
+			sb << "Node " << node << ", setting output to {";
+			for ( auto i = 0u ; i < sInput.set.size( ) ; i ++ ) {
+				if ( i ) {
+					sb << " ,";
+				}
+				sb << ' ' << sInput.set[ i ];
+			}
+			sb << " }";
+			return sb;
+		} , LL2 );
+#endif
+		sOut = sInput;
+
+		// Add the node's successors into the changed set
+		const auto nOut{ cn.outbound.size( ) };
+		for ( auto i = 0u ; i < nOut ; i ++ ) {
+			auto const& ob{ cn.outbound[ i ] };
+			// Make sure the outbound link type is right (FLOW
+			// always is, BYPASS works in functions, CALL/RET at
+			// the global level)
+			if ( ob.type != ob.FLOW && (
+					( fnIndex && ob.type != ob.BYPASS )
+					|| ( !fnIndex && ob.type == ob.BYPASS ) ) ) {
+				continue;
+			}
+			changed.add( ob.target );
+		}
+	}
+
+	// Now we need to identify the blocks which contain var uses
+	T_Set< uint32_t > blocks{ UseTag< ArrayBacked< 16 > >( ) }; // FIXME alloc
+	for ( auto i = 0u ; i < var.uses.size( ) ; i ++ ) {
+		const auto useInstr{ var.uses[ i ].node };
+		for ( auto j = nFirst ; j < nEnd ; j ++ ) {
+			auto const& block{ *data.ctrlFlowGraph[ j ] };
+			if ( !block.instructions ) {
+				continue;
+			}
+			const auto f{ block.instructions->first };
+			const auto e{ block.instructions->count + f - 1 };
+			if ( f <= useInstr && e >= useInstr ) {
+				blocks.add( j );
+				break;
+			}
+		}
+	}
+
+	// Now go through all the blocks instruction by instruction, updating
+	// the active definition set if a definition is found, and associating
+	// defines to uses.
+	for ( auto i = 0u ; i < blocks.size( ) ; i ++ ) {
+		const auto bid{ blocks[ i ] };
+		auto const& cb{ *data.ctrlFlowGraph[ bid ] };
+		T_BUDCDefSet_ defs{ defines[ bid ] };
+		assert( cb.instructions );
+
+		const auto is{ cb.instructions->first };
+		const auto ie{ is + cb.instructions->count };
+		for ( auto ii = is ; ii < ie ; ii ++ ) {
+			auto const* const irec{ udPerInstr.get( ii ) };
+			if ( !irec ) {
+				continue;
+			}
+			const auto nrec{ irec->size( ) };
+
+			// Handle uses first
+			for ( auto j = 0u ; j < nrec ; j ++ ) {
+				auto const& rec{ (*irec)[ j ] };
+				auto const& ud{ data.varUDChains[ rec.entry ] };
+				if ( !rec.isUse || ud.var != var.var ) {
+					continue;
+				}
+				for ( auto di = 0u ; di < defs.set.size( ) ; di ++ ) {
+					BUDCLink_( var , defs.set[ di ] ,
+							rec.index , data.logger );
+				}
+				if ( !defs.set.size( ) ) {
+					// FIXME: add warning
+					BUDCLink_( var , T_HashIndex::INVALID_INDEX ,
+							rec.index , data.logger );
+				}
+			}
+
+			// Update the set
+			for ( auto j = 0u ; j < nrec ; j ++ ) {
+				auto const& rec{ (*irec)[ j ] };
+				auto const& ud{ data.varUDChains[ rec.entry ] };
+				if ( rec.isUse || ud.var != var.var ) {
+					continue;
+				}
+				defs.clear( );
+				defs.add( rec.index );
+				break;
+			}
+		}
+	}
+}
+
 
 } // namespace <anon>
 
@@ -828,7 +1087,7 @@ void T_OptData::buildUseDefineChains(
 	for ( auto& sym : varUDChains.values( ) ) {
 		switch ( sym.var.type ) {
 		    case T_OptData::E_UDVarType::GLOBAL:
-#warning TODO
+			BUDCWalkGraph_( *this , sym , udPerInstr , {} );
 			break;
 
 		    case T_OptData::E_UDVarType::ARGUMENT: {
@@ -842,9 +1101,12 @@ void T_OptData::buildUseDefineChains(
 			break;
 		    }
 
-		    case T_OptData::E_UDVarType::LOCAL:
+		    case T_OptData::E_UDVarType::LOCAL: {
-#warning TODO
+			assert( cfgFunctions.contains( sym.var.owner ) );
+			BUDCWalkGraph_( *this , sym , udPerInstr ,
+					cfgFunctions.indexOf( sym.var.owner ) );
 			break;
+		    }
 		}
 	}