MimIR
MimIR is my Intermediate Representation
Loading...
Searching...
No Matches
mim::plug::tensor::phase::LowerToMem Class Reference

Bufferizes the low-level tensor axioms onto the shared buffer layer. More...

#include <mim/plug/tensor/phase/lower_to_mem.h>

Inheritance diagram for mim::plug::tensor::phase::LowerToMem:
[legend]

Public Member Functions

 LowerToMem (World &world, flags_t annex)
Public Member Functions inherited from mim::RWPhase
 RWPhase (World &world, std::string name, Analysis *analysis=nullptr)
 RWPhase (World &world, flags_t annex, Analysis *analysis=nullptr)
void rewrite_annex (flags_t, Sym, const Def *) override
void rewrite_external (Def *) override
Worldworld ()=delete
 Hides both and forbids direct access.
Worldold_world ()
 Get old Defs from here.
Worldnew_world ()
 Create new Defs into this.
Public Member Functions inherited from mim::RWBase
Analysisanalysis ()
const Analysisanalysis () const
const Deflattice (const Def *def) const
 Returns the abstract value computed by the associated Analysis for def, or nullptr if no value is available.
const Defabstracted (const Def *def) const
 Returns lattice(def) if it differs from def (i.e. we learned something), otherwise nullptr.
virtual bool analyze ()
 Runs the optional pre-analysis on Phase::world, typically to a fixed point, before rewriting begins.
bool is_bootstrapping () const
 Returns whether we are currently bootstrapping (rewriting annexes).
Public Member Functions inherited from mim::Phase
 Phase (World &world, std::string name)
 Phase (World &world, flags_t annex)
virtual ~Phase ()=default
virtual std::unique_ptr< Phaserecreate ()
 Creates a new instance; needed by a fixed-point PhaseMan.
virtual void apply (const App *)
 Invoked if your Phase has additional args.
virtual void apply (Phase &)
 Dito, but invoked by Phase::recreate.
virtual bool redirects () const
 If true, Phase::create uses take_resolved().
virtual std::unique_ptr< Phasetake_resolved ()
 The Phase to use instead; nullptr means elide.
Worldworld ()
Driverdriver ()
const fe::Log & log () const
std::string_view name () const
flags_t annex () const
const fe::Vector< std::string > & args ()
 Command-line arguments passed to this Phase's plugin via -X <plugin>:<arg>.
bool todo () const
void invalidate (bool todo=true)
 Signals that another round of fixed-point iteration is required, either as part of.
virtual void run ()
 Entry point and generates some debug output; invokes Phase::start.
void profile_count (std::string_view key, uint64_t n=1)
 Adds n to the custom fe::Profiler counter key of the current run; no-op unless profiling is enabled.
Public Member Functions inherited from mim::Rewriter
template<class D = Def>
D * curr_mut () const
 Rewriter (std::unique_ptr< World > &&ptr)
 Rewriter (World &world)
virtual ~Rewriter ()
void reset (std::unique_ptr< World > &&ptr)
void reset ()
Worldworld ()
virtual void push ()
virtual void pop ()
virtual const Defmap (const Def *old_def, const Def *new_def)
const Defmap_root (const Def *old_def, const Def *new_def)
 Like map() but records into the root map, so the entry outlives the current push()/pop() scope.
const Defmap (const Def *old_def, Defs new_defs)
const Defmap (Defs old_defs, const Def *new_def)
const Defmap (Defs old_defs, Defs new_defs)
virtual const Deflookup (const Def *old_def)
 Lookup old_def by searching in reverse through the stack of maps.
virtual const Defrewrite_imm (const Def *)
virtual const Defrewrite_mut (Def *)
virtual const Defrewrite_stub (Def *, Def *)
virtual DefVec rewrite (Defs)
virtual const Defrewrite_imm_Seq (const Seq *seq)
virtual const Defrewrite_mut_Seq (Seq *seq)

Private Member Functions

void start () override
 Actual entry.
const Defrewrite (const Def *) override
const Defrewrite_mut_Lam (Lam *) override
const Defrewrite_imm_App (const App *) override

Additional Inherited Members

Static Public Member Functions inherited from mim::Phase
static std::unique_ptr< Phasecreate (const Flags2Phases &phases, const Def *def)
template<class A, class P>
static void hook (Flags2Phases &phases)
template<class P, class... Args>
static void run (Args &&... args)
 Runs a single Phase.
Protected Member Functions inherited from mim::RWPhase
void start () override
 RWBase::start() and then swaps the two worlds.
Protected Member Functions inherited from mim::RWBase
virtual const Defrewrite_root (const Def *def)
 Rewrites a root - i.e. an annex or an external.
virtual void finalize ()
 Run after all roots have been walked - but for an RWPhase still before the two worlds are swapped.
 RWBase (World &world, std::string name, Analysis *analysis)
 RWBase (World &world, flags_t annex, Analysis *analysis)
 RWBase (World &world, std::string name, Analysis *analysis, std::unique_ptr< World > &&new_world)
 Rewrite the World of Phase::world into new_world.
 RWBase (World &world, flags_t annex, Analysis *analysis, std::unique_ptr< World > &&new_world)
Protected Member Functions inherited from mim::Rewriter
auto enter (Def *new_mut)
 Updates curr_mut() to new_mut and restores it at the end of the scope.
Protected Attributes inherited from mim::Phase
std::string name_
Protected Attributes inherited from mim::Rewriter
std::deque< Def2Defold2news_

Detailed Description

Bufferizes the low-level tensor axioms onto the shared buffer layer.

get / set become buffer.read / buffer.write, generate becomes an allocating fill loop, map_reduce / broadcast / pad / concat / gather / scatter become their buffer-world btensor.* counterparts, and tensor array values «s; T» become buffer.Buf (r, s, T) handles. Afterwards buffer.lower_ptr lowers the buffer layer to mem.Ptr + mem.lea / mem.load / mem.store.

This phase is conversion-only: it rewrites types and operations but does not thread the mem.M memory monad itself. Emitted buffer operations consume a ⊥: mem.M 0 placeholder (or a short local chain rooted in a LowerToMem::fresh_mem continuation's var), and the SSA value dependencies keep them anchored and ordered. The mem.add_mem phase (mim::plug::mem::phase::AddMem), scheduled right after this one in the pipeline, then mem-extends all continuations and rewires every memory operand to the scheduler-placed current memory — handling returns, error continuations, join points, branch arms, and interleaving with a caller's own memory operations uniformly.

Which array types denote tensors (as opposed to index/shape arrays that share the Arr structure) is determined by role: a pre-pass collects the array operand/result types of the tensor operations, and only those types are rewritten to Buf — and only at function boundaries, never as a global type rewrite.

Definition at line 28 of file lower_to_mem.h.

Constructor & Destructor Documentation

◆ LowerToMem()

mim::plug::tensor::phase::LowerToMem::LowerToMem ( World & world,
flags_t annex )
inline

Member Function Documentation

◆ rewrite()

const Def * mim::plug::tensor::phase::LowerToMem::rewrite ( const Def * old_def)
overrideprivatevirtual

Reimplemented from mim::Rewriter.

Definition at line 306 of file lower_to_mem.cpp.

References mim::RWBase::is_bootstrapping(), mim::Rewriter::rewrite(), and rewrite_imm_App().

Referenced by rewrite_imm_App().

◆ rewrite_imm_App()

const Def * mim::plug::tensor::phase::LowerToMem::rewrite_imm_App ( const App * app)
overrideprivate

◆ rewrite_mut_Lam()

const Def * mim::plug::tensor::phase::LowerToMem::rewrite_mut_Lam ( Lam * lam)
overrideprivate

Definition at line 348 of file lower_to_mem.cpp.

References mim::RWBase::is_bootstrapping().

◆ start()

void mim::plug::tensor::phase::LowerToMem::start ( )
overrideprivatevirtual

Actual entry.

Implements mim::Phase.

Definition at line 253 of file lower_to_mem.cpp.

References mim::RWPhase::start().


The documentation for this class was generated from the following files: