Merge POV bits where a single POV bit could be shared between multiple headers. More...

#include <merge_pov_bits.h>

Inheritance diagram for BFN::MergePovBits:

Public Member Functions
	MergePovBits (const PhvInfo &phv)

Public Member Functions inherited from P4::PassManager
	PassManager (const PassManager &)=default

	PassManager (const std::initializer_list< VisitorRef > &init)

	PassManager (PassManager &&)=default

void	addDebugHook (DebugHook h, bool recursive=false)

void	addDebugHooks (std::vector< DebugHook > hooks, bool recursive=false)

void	addPasses (const std::initializer_list< VisitorRef > &init)

const IR::Node *	apply_visitor (const IR::Node , const char =0) override

bool	backtrack (trigger &trig) override

PassManager *	clone () const override

void	early_exit ()

void	listPasses (std::ostream &, cstring sep) const

bool	never_backtracks () override

void	removePasses (const std::vector< cstring > &exclude)

void	setStopOnError (bool stop)

Public Member Functions inherited from P4::Visitor
virtual bool	check_clone (const Visitor *a)

virtual bool	check_global (cstring)

virtual void	clear_globals ()

virtual ControlFlowVisitor *	controlFlowVisitor ()

virtual void	end_apply ()

virtual void	end_apply (const IR::Node *root)

virtual void	erase_global (cstring)

template<class T >
const T *	findContext () const

template<class T >
const T *	findContext (const Context *&c) const

template<class T >
const T *	findOrigCtxt () const

template<class T >
const T *	findOrigCtxt (const Context *&c) const

virtual Visitor &	flow_clone ()

virtual void	flow_merge (Visitor &)

virtual bool	flow_merge_closure (Visitor &)

virtual void	flow_merge_global_from (cstring)

virtual void	flow_merge_global_to (cstring)

const Context *	getChildContext () const

int	getChildrenVisited () const

const Context *	getContext () const

int	getContextDepth () const

const IR::Node *	getCurrentNode () const

template<class T >
const T *	getCurrentNode () const

const IR::Node *	getOriginal () const

template<class T >
const T *	getOriginal () const

template<class T >
const T *	getParent () const

virtual bool	has_flow_joins () const

profile_t	init_apply (const IR::Node root, const Context parent_context)

bool	isInContext (const IR::Node *n) const

virtual const char *	name () const

template<class T >
void	parallel_visit (const IR::Vector< T > &v, const char *name, int cidx)

template<class T >
void	parallel_visit (const IR::Vector< T > &v, const char *name=0)

template<class T >
void	parallel_visit (IR::Vector< T > &v, const char *name, int cidx)

template<class T >
void	parallel_visit (IR::Vector< T > &v, const char *name=0)

void	print_context () const

const Visitor &	setCalledBy (const Visitor *visitor)

void	setName (const char *name)

void	visit (const IR::Node &n, const char *name, int cidx)

void	visit (const IR::Node &n, const char *name=0)

void	visit (const IR::Node &n, const char name, int cidx)

void	visit (const IR::Node &n, const char name=0)

void	visit (const IR::Node const &n, const char name, int cidx)

void	visit (const IR::Node const &n, const char name=0)

void	visit (IR::Node &n, const char *name, int cidx)

void	visit (IR::Node &n, const char *name=0)

void	visit (IR::Node &, const char =0, int=0)

template<class T , typename = std::enable_if_t<Util::has_SourceInfo_v<T> && !std::is_pointer_v<T>>, class... Args>
void	warn (const int kind, const char *format, const T &node, Args &&...args)
	The const ref variant of the above.

template<class T , typename = std::enable_if_t<Util::has_SourceInfo_v<T>>, class... Args>
void	warn (const int kind, const char format, const T node, Args &&...args)

bool	warning_enabled (int warning_kind) const

Public Member Functions inherited from P4::Backtrack

Protected Attributes
ordered_map< const PHV::Field , const PHV::Field >	merge_pov

Protected Attributes inherited from P4::PassManager
safe_vector< DebugHook >	debugHooks

safe_vector< Visitor * >	passes

bool	running = false

unsigned	seqNo = 0

bool	stop_on_error = true

Protected Attributes inherited from P4::Visitor
bool	dontForwardChildrenBeforePreorder = false

bool	joinFlows = false

bool	visitDagOnce = true

Additional Inherited Members
Public Types inherited from P4::Visitor
typedef Visitor_Context	Context

Static Public Member Functions inherited from P4::Visitor
static cstring	demangle (const char *)

static bool	warning_enabled (const Visitor *visitor, int warning_kind)

Public Attributes inherited from P4::Visitor
const Visitor *	called_by = nullptr

cstring	internalName

SplitFlowVisit_base *&	split_link

SplitFlowVisit_base *	split_link_mem = nullptr

Protected Member Functions inherited from P4::PassManager
profile_t	init_apply (const IR::Node *root) override

void	runDebugHooks (const char visitorName, const IR::Node node)

Protected Member Functions inherited from P4::Visitor
virtual void	init_join_flows (const IR::Node *)

virtual bool	join_flows (const IR::Node *)

virtual void	post_join_flows (const IR::Node , const IR::Node )

void	visit_children (const IR::Node *, std::function< void()> fn)

virtual void	visitAgain () const

virtual void	visitOnce () const

Detailed Description

Merge POV bits where a single POV bit could be shared between multiple headers.

Identify situations where a single POV bit could be shared across multiple headers. "Merge" the POV bits by replacing with the single POV bit target in the deparser.

This pass is currently fairly simple: it only replaces references in the deparser. Ideally it would replace all references as this would save some POV bits and could help with PHV fitting.

I had experimented with more complete POV replacement throughout the pipe but halted those changes. There are several places where more extensive changes are required for a more full replacement:

Header mutex analysis: Logic doesn't account for aliased POV bits, so a manipulation of an alias source doesn't register as updating the alias destination and vice-versa.
Usedef: Logic doesn't account for aliased POV bits.
IXBar + hash: If two POV bits that are merged are used in a hash, the bits still need to be included twice. Currently the compiler will duplicate the byte in the IXBar (once for source and once for dest), but the assembler doesn't currently handle this correctly. Can either allow the byte duplication (consuming more IXBar resources) and update the assembler, or eliminate the duplication in the compiler but this requires allowing duplicate bit refences in IXBar::Use::Byte objects.
IXBar + match: Requires either the byte duplication on the ixbar, or duplication of bytes on the match bus. (Either choice consumes additional resources of one type or another.)

The current approach of only replacing the POV in the deparser avoids the problems listed above. However, it must be run after code elimination, otherwise set valid/set invalid calls may be eliminated and then header mutex analysis breaks.

Public Member Functions

Protected Attributes

Additional Inherited Members

Detailed Description