The following are various notes about the design and implementation of LuaJIT.
Design overview
Email to lua-l mailing list.
From: Mike Pall mike.de
Subject: LuaJIT 2.0 intellectual property disclosure and research opportunities
Newsgroups: gmane.comp.lang.lua.general
Date: Monday 2nd November 2009 10:17:04 UTC (over 7 years ago)
It has been brought to my attention that it might be advantageous
for some parts of the research community and the open source
community, that I make a public statement about the intellectual
property (IP) contained in LuaJIT 2.0 and earlier versions:
I hereby declare any and all of my own inventions contained in
LuaJIT to be in the public domain and up for free use by anyone
without payment of any royalties whatsoever.
[Note that the source code itself is licensed under a permissive
license and is not placed in the public domain. But this is an
orthogonal issue.]
I cannot guarantee it to be free of third-party IP however. In
fact nobody can. Writing software has become a minefield and any
moderately complex piece of software is probably (unknowingly to
the author) encumbered by hundreds of dubious patents. This
especially applies to compilers. The curent IP system is broken
and software patents must be abolished. Ceterum censeo.
The usual form of disclosure is to write papers and publish them.
I'm sorry, but I don't have the time for this right now. But I
would consider publishing open source software as a form of
disclosure.
In the interest of anyone doing research on virtual machines,
compilers and interpreters, I've compiled a list of some of the
new aspects to be found in LuaJIT 2.0. I do not claim all of them
are original (I cannot possibly know all of the literature), but
my research indicates that many of them are quite innovative.
This also presents some research opportunities for 3rd parties.
I have little use for academic merits myself -- I'm more interested
in coding than writing papers. Anyone is welcome to dig out any
aspects, explore them in detail and publish them (giving due credit).
Design aspects of the VM:
- NaN-tagging: 64 bit tagged values are used for stack slots and
table slots. Unboxed floating-point numbers (doubles) are
overlayed with tagged object references. The latter can be
distinguished from numbers via the use of special NaNs as tags.
It's a remote descendant of pointer-tagging.
[The idea dates back to 2006, but I haven't disclosed it before
2008. Special NaNs have been used to overlay pointers before.
Others have used it for tagging later on. The specific layout is
of my own devising.]
- Low-overhead call frames: The linear, growable stack implicitly
holds the frame structure. The tags for the base function of
each call frame hold a linked structure of frames, using no
extra space. Calls/returns are faster due to lower memory
traffic. This also allows installing exception handlers at zero
cost (it's a special bit pattern in the frame link).
Design of the IR (intermediate representation) used by the compiler:
- Linear, pointer-free IR: The typed IR is SSA-based and highly
orthogonal. An instruction takes up only 64 bits. It has up to
two operands which are 16 bit references. It's implemented with
a bidirectionally growable array. No trees, no pointers, no cry.
Heavily optimized for minimal D-cache impact, too.
- Skip-list chains: The IR is threaded with segregated, per-opcode
skip-list chains. The links are stored in a multi-purpose 16 bit
field in the instruction. This facilitates low-overhead lookup
for CSE, DSE and alias analysis. Back-linking enables short-cut
searches (average overhead is less than 1 lookup). Incremental
build-up is trivial. No hashes, no sets, no complex updates.
- IR references: Specially crafted IR references allow fast const
vs. non-const decisions. The trace recorder uses type-tagged
references (a form of caching) internally for low-overhead
type-based dispatch.
- High-level IR: A single, uniform high-level IR is used across
all stages of the compiler. This reduces overall complexity.
Careful instruction design avoids any impact on low-level CSE
opportunities. It also allows cheap and effective high-level
semantic disambiguation for memory references.
Design of the compiler pipeline:
- Rule-based FOLD engine: The FOLD engine is primarily used for
constant folding, algebraic simplifications and reassociation.
Most traditional compilers have an evolutionary grown set of
implicit rules, spread over thousands of hand-coded tiny
conditionals.
The rule-based FOLD engine uses a declarative approach to
combine the first and second level of lookup. It allows wildcard
lookup with masked keys, too. A pre-processor generates a
semi-perfect hash table for constant-time rule lookup. It's able
to deal with thousands of rules in a uniform manner without
performance degradation. A declarative approach is also much
easier to maintain.
- Unified stage dispatch: The FOLD engine is the first stage in
the compiler pipeline. Wildcard rules are used to dispatch
specific instructions or instruction types (loads, stores,
allocations etc.) to later optimization stages (load forwarding,
DSE etc.). Unmatched instructions are passed on to CSE.
Unified stage dispatch facilitates modular and pluggable
optimizations with only local knowledge. It's also faster than
doing multiple dispatches in every stage.
Trace compiler:
- NLF region-selection: The trace heuristics use a natural-loop
first (NLF) region-selection mechanism to come up with a
close-to optimal set of (looping) root traces. Only special
bytecode instructions trigger new root traces -- regular
conditionals never do this. Root traces that leave the loop are
aborted and retried later. This also gives outer loops a chance
to inline inner loops with a low trip count.
NLF usually generates a superior set of root traces than the
MRET/NET (next-executing tail) and LEI (last-executed iteration)
region-selection mechanisms known from the literature.
- Hashed profile counters: Bytecode instructions to trigger the
start of a hot trace use low-overhead hashed profiling counters.
The profile is imprecise because collisions are ignored. The
hash table is kept very small to reduce D-cache impact (only two
hot cache lines). Since NLF weeds out most false positives, this
doesn't deteriorate hot trace detection.
[Neither using hashed profile counters, nor imprecise profiling,
nor using profiling to detect hot loops is new. But the specific
combination may be original.]
- Code sinking via snapshots: The VM must be in a consistent state
when a trace exits. This means that all updates (stores) to the
state (stack or objects) must track the original language
semantics.
Naive trace compilers achieve this by forcing a full update of
the state to memory before every exit. This causes many on-trace
stores and seriously diminishes code quality.
A better approach is to sink these stores to compensation code,
which is only executed if the trace exits are actually taken.
A common solution is to emit actual code for these stores. But
this causes code cache bloat and the information often needs to
be stored redundantly, for linking of side traces.
Code sinking via snapshots allows sinking of arbitrary code
without the overhead of the other approaches. A snapshot stores
a consistent view of all updates to the state before an exit. If
an exit is taken the on-trace machine state (registers and spill
slots) and the snapshot can be used to restore the VM state.
State restoration using this data-driven approach is slow of
course. But repeatedly taken side exits quickly trigger the
generation of side traces. The snapshot is used to initialize
the IR of the side trace with the necessary state using
pseudo-loads. These can be optimized together with the remainder
of the side trace. The pseudo-loads are unified with the machine
state of the parent trace by the backend to enable zero-cost
linking to side traces.
[Currently snapshots only allow store sinking of scalars. It's
planned to extend this to allow arbitrary store and allocation
sinking, which together with store forwarding would be a unique
way to achieve scalar-replacement of aggregates.]
- Sparse snapshots: Taking a full snapshot of all state updates
before every exit would need a considerable amount of storage.
Since all scalar stores are sunk, it's feasible to reduce the
snapshot density. The basic idea is that it doesn't matter which
state is restored on a taken exit, as long as it's consistent.
This is a form of transactional state management. Every snapshot
is a commit; a taken exit causes a rollback to the last commit.
The on-trace state may advance beyond the last commit as long as
this doesn't affect the possibility of a rollback. In practice
this means that all on-trace updates to the state (non-scalar
stores that are not sunk) need to force a new snapshot for the
next exit.
Otherwise the trace recorder only generates a snapshot after
control-flow constructs that are present in the source, too.
Guards that have a low probability of being wrongly predicted do
not cause snapshots (e.g. function dispatch). This further
reduces the snapshot density. Sparse snapshots also improve
on-trace code quality, because they reduce the live range of the
results of intermediate computations. Scheduling decisions can
be made over a longer stream of instructions, too.
[It's planned to switch to compressed snapshots. 2D-compression
across snapshots may be able to remove even more redundancy.]
Optimizations:
- Hash slot specialization: Hash table lookup for constant keys is
specialized to the predicted hash slot. This avoids a loop to
follow the hash chain. Pseudocode:
HREFK: if (hash[17].key != key) goto exit
HLOAD: x = hash[17].value
-or-
HSTORE: hash[17].value = x
HREFK is shared by multiple HLOADs/HSTOREs and may be hoisted
independently. The verification of the prediction (HREFK) is
moved out of the dependency chain by a super-scalar CPU. This
makes hash lookup as cheap as array lookup with minimal complexity.
It also avoids all the complications (cache invalidation,
ordering constraints, shape mismatches) associated with hidden
classes (V8) or shape inference/property caching (TraceMonkey).
- Code hoisting via unrolling and copy-substitution (LOOP):
Traditional loop-invariant code motion (LICM) is mostly useless
for the IR resulting from dynamic languages. The IR has many
guards and most subsequent instructions are control-dependent on
them. The first non-hoistable guard would effectively prevent
hoisting of all subsequent instructions.
The LOOP pass does synthetic unrolling of the recorded IR,
combining copy-substitution with redundancy elimination to
achieve code hoisting. The unrolled and copy-substituted
instructions are simply fed back into the compiler pipeline,
which allows reuse of all optimizations for redundancy
elimination. Loop recurrences are detected on-the-fly and a
minimized set of PHIs is generated.
- Narrowing of numbers to integers: Predictive narrowing is used
for induction variables. Demand-driven narrowing is used for
index expressions using a backpropagation algorithm.
This avoids the complexity associated with speculative, eager
narrowing, which also causes excessive control-flow dependencies
due to the many overflow checks. Selective narrowing is better
at exploiting the combined bandwidth of the FP and integer units
of the CPU and avoids clogging up the branch unit.
Register allocation:
- Blended cost-model for R-LSRA: The reverse-linear-scan register
allocator uses a blended cost model for its spill decisions.
This takes into account multiple factors (e.g. PHI weight) and
benefits from the special layout of IR references (constants
before invariant instructions, before variant instructions).
- Register hints: The register allocation heuristics take into
account register hints, e.g. for loop recurrences or calling
conventions. This is very cheap to implement, but improves the
allocation decisions considerably. It reduces register shuffling
and prevents unnecessary spills.
- x86-specific improvements: Special heuristics for move vs.
rename produce close to optimal code for two-operand machine
code instructions.
Fusion of memory operands into instructions is required to
generate high-quality x86 code. Late fusion in the backend
allows better, local decisions, based on actual register
pressure, rather than estimates of prior stages.
Ok, that's it! Sorry for the length of this posting, but I hope it
was at least informative to someone out there.
--Mike
Links
