with the new pass manager, and no longer relying on analysis groups.
This builds essentially a ground-up new AA infrastructure stack for
LLVM. The core ideas are the same that are used throughout the new pass
manager: type erased polymorphism and direct composition. The design is
as follows:
- FunctionAAResults is a type-erasing alias analysis results aggregation
interface to walk a single query across a range of results from
different alias analyses. Currently this is function-specific as we
always assume that aliasing queries are *within* a function.
- AAResultBase is a CRTP utility providing stub implementations of
various parts of the alias analysis result concept, notably in several
cases in terms of other more general parts of the interface. This can
be used to implement only a narrow part of the interface rather than
the entire interface. This isn't really ideal, this logic should be
hoisted into FunctionAAResults as currently it will cause
a significant amount of redundant work, but it faithfully models the
behavior of the prior infrastructure.
- All the alias analysis passes are ported to be wrapper passes for the
legacy PM and new-style analysis passes for the new PM with a shared
result object. In some cases (most notably CFL), this is an extremely
naive approach that we should revisit when we can specialize for the
new pass manager.
- BasicAA has been restructured to reflect that it is much more
fundamentally a function analysis because it uses dominator trees and
loop info that need to be constructed for each function.
All of the references to getting alias analysis results have been
updated to use the new aggregation interface. All the preservation and
other pass management code has been updated accordingly.
The way the FunctionAAResultsWrapperPass works is to detect the
available alias analyses when run, and add them to the results object.
This means that we should be able to continue to respect when various
passes are added to the pipeline, for example adding CFL or adding TBAA
passes should just cause their results to be available and to get folded
into this. The exception to this rule is BasicAA which really needs to
be a function pass due to using dominator trees and loop info. As
a consequence, the FunctionAAResultsWrapperPass directly depends on
BasicAA and always includes it in the aggregation.
This has significant implications for preserving analyses. Generally,
most passes shouldn't bother preserving FunctionAAResultsWrapperPass
because rebuilding the results just updates the set of known AA passes.
The exception to this rule are LoopPass instances which need to preserve
all the function analyses that the loop pass manager will end up
needing. This means preserving both BasicAAWrapperPass and the
aggregating FunctionAAResultsWrapperPass.
Now, when preserving an alias analysis, you do so by directly preserving
that analysis. This is only necessary for non-immutable-pass-provided
alias analyses though, and there are only three of interest: BasicAA,
GlobalsAA (formerly GlobalsModRef), and SCEVAA. Usually BasicAA is
preserved when needed because it (like DominatorTree and LoopInfo) is
marked as a CFG-only pass. I've expanded GlobalsAA into the preserved
set everywhere we previously were preserving all of AliasAnalysis, and
I've added SCEVAA in the intersection of that with where we preserve
SCEV itself.
One significant challenge to all of this is that the CGSCC passes were
actually using the alias analysis implementations by taking advantage of
a pretty amazing set of loop holes in the old pass manager's analysis
management code which allowed analysis groups to slide through in many
cases. Moving away from analysis groups makes this problem much more
obvious. To fix it, I've leveraged the flexibility the design of the new
PM components provides to just directly construct the relevant alias
analyses for the relevant functions in the IPO passes that need them.
This is a bit hacky, but should go away with the new pass manager, and
is already in many ways cleaner than the prior state.
Another significant challenge is that various facilities of the old
alias analysis infrastructure just don't fit any more. The most
significant of these is the alias analysis 'counter' pass. That pass
relied on the ability to snoop on AA queries at different points in the
analysis group chain. Instead, I'm planning to build printing
functionality directly into the aggregation layer. I've not included
that in this patch merely to keep it smaller.
Note that all of this needs a nearly complete rewrite of the AA
documentation. I'm planning to do that, but I'd like to make sure the
new design settles, and to flesh out a bit more of what it looks like in
the new pass manager first.
Differential Revision: http://reviews.llvm.org/D12080
llvm-svn: 247167
PR24139 contains an analysis of poor register allocation. One of the findings
was that when calculating the spill weight, a rematerializable interval once
split is no longer rematerializable. This is because the isRematerializable
check in CalcSpillWeights.cpp does not follow the copies introduced by live
range splitting (after splitting, the live interval register definition is a
copy which is not rematerializable).
Reviewers: qcolombet
Differential Revision: http://reviews.llvm.org/D11686
llvm-svn: 244439
Build time (user time) for building llvm+clang+lldb in release mode:
- default allocator: 9086 seconds
- with PBQP: 9126 seconds
- with PBQP + local bit matrix cache: 9097 seconds
llvm-svn: 231360
Such edges are zero matrix, and they bring no additional info to the
allocation problem, apart from contributing to nodes' degree. Removing
those edges is expected to improve allocation time.
Tune the spill cost comparison, as this gives better average performances
now that the nodes' degrees has changed.
llvm-svn: 230904
Although such nodes are allocatable, the cost of spilling may be less than
allocating to register, so spilling the node may provide a better solution.
The assert does not account for this case, so remove it for now.
llvm-svn: 229103
The NodeMetadata are maintained in an incremental way. When an edge between
2 nodes has its cost updated, in the course of graph reduction for example,
the NodeMetadata need first to have the old edge cost removed, then the new
edge cost added. Only once the NodeMetadata have been fully updated, it
becomes safe to consider promoting the nodes to the
ConservativelyAllocatable or OptimallyReducible sets. Previously, this
promotion was occuring right after the removing the old cost, and this was
breaking the assumption that a ConservativelyAllocatable should not be
spilled.
This patch also adds asserts to:
- enforces the invariant that a node's reduction can not be downgraded,
- only not provably allocatable or optimally reducible nodes can be spilled.
llvm-svn: 228816
The PBQP::RegAlloc::MatrixMetadata class assumes that matrices have at least two
rows/columns (for the spill option plus at least one physreg). This patch
ensures that that invariant is met by pre-spilling vregs that have no physreg
options so that no node (and no corresponding edges) need be added to the PBQP
graph.
This fixes a bug in an out-of-tree target that was identified by Jonas Paulsson.
Thanks for tracking this down Jonas!
llvm-svn: 227942
Registers are not all equal. Some are not allocatable (infinite cost),
some have to be preserved but can be used, and some others are just free
to use.
Ensure there is a cost hierarchy reflecting this fact, so that the
allocator will favor scratch registers over callee-saved registers.
llvm-svn: 221293
This patch improves how the different costs (register, interference, spill
and coalescing) relates together. The assumption is now that:
- coalescing (or any other "side effect" of reg alloc) is negative, and
instead of being derived from a spill cost, they use the block
frequency info.
- spill costs are in the [MinSpillCost:+inf( range
- register or interference costs are in [0.0:MinSpillCost( or +inf
The current MinSpillCost is set to 10.0, which is a random value high
enough that the current constraint builders do not need to worry about
when settings costs. It would however be worth adding a normalization
step for register and interference costs as the last step in the
constraint builder chain to ensure they are not greater than SpillMinCost
(unless this has some sense for some architectures). This would work well
with the current builder pipeline, where all costs are tweaked relatively
to each others, but could grow above MinSpillCost if the pipeline is
deep enough.
The current heuristic is tuned to depend rather on the number of uses of
a live interval rather than a density of uses, as used by the greedy
allocator. This heuristic provides a few percent improvement on a number
of benchmarks (eembc, spec, ...) and will definitely need to change once
spill placement is implemented: the current spill placement is really
ineficient, so making the cost proportionnal to the number of use is a
clear win.
llvm-svn: 221292
sets as keys into a cache of interference matrice values in the Interference
constraint adder.
Creating interference matrices was one of the large remaining time-sinks in
PBQP. Caching them reduces the total compile time (when using PBQP) on the
nightly test suite by ~10%.
llvm-svn: 220688
This patch removes the PBQPBuilder class and its subclasses and replaces them
with a composable constraints class: PBQPRAConstraint. This allows constraints
that are only required for optimisation (e.g. coalescing, soft pairing) to be
mixed and matched.
This patch also introduces support for target writers to supply custom
constraints for their targets by overriding a TargetSubtargetInfo method:
std::unique_ptr<PBQPRAConstraints> getCustomPBQPConstraints() const;
This patch should have no effect on allocations.
llvm-svn: 219421
Take a StringRef instead of a "const char *".
Take a "std::error_code &" instead of a "std::string &" for error.
A create static method would be even better, but this patch is already a bit too
big.
llvm-svn: 216393
shorter/easier and have the DAG use that to do the same lookup. This
can be used in the future for TargetMachine based caching lookups from
the MachineFunction easily.
Update the MIPS subtarget switching machinery to update this pointer
at the same time it runs.
llvm-svn: 214838
define below all header includes in the lib/CodeGen/... tree. While the
current modules implementation doesn't check for this kind of ODR
violation yet, it is likely to grow support for it in the future. It
also removes one layer of macro pollution across all the included
headers.
Other sub-trees will follow.
llvm-svn: 206837
This compiles with no changes to clang/lld/lldb with MSVC and includes
overloads to various functions which are used by those projects and llvm
which have OwningPtr's as parameters. This should allow out of tree
projects some time to move. There are also no changes to libs/Target,
which should help out of tree targets have time to move, if necessary.
llvm-svn: 203083
The previous PBQP solver was very robust but consumed a lot of memory,
performed a lot of redundant computation, and contained some unnecessarily tight
coupling that prevented experimentation with novel solution techniques. This new
solver is an attempt to address these shortcomings.
Important/interesting changes:
1) The domain-independent PBQP solver class, HeuristicSolverImpl, is gone.
It is replaced by a register allocation specific solver, PBQP::RegAlloc::Solver
(see RegAllocSolver.h).
The optimal reduction rules and the backpropagation algorithm have been extracted
into stand-alone functions (see ReductionRules.h), which can be used to build
domain specific PBQP solvers. This provides many more opportunities for
domain-specific knowledge to inform the PBQP solvers' decisions. In theory this
should allow us to generate better solutions. In practice, we can at least test
out ideas now.
As a side benefit, I believe the new solver is more readable than the old one.
2) The solver type is now a template parameter of the PBQP graph.
This allows the graph to notify the solver of any modifications made (e.g. by
domain independent rules) without the overhead of a virtual call. It also allows
the solver to supply policy information to the graph (see below).
3) Significantly reduced memory overhead.
Memory management policy is now an explicit property of the PBQP graph (via
the CostAllocator typedef on the graph's solver template argument). Because PBQP
graphs for register allocation tend to contain many redundant instances of
single values (E.g. the value representing an interference constraint between
GPRs), the new RASolver class uses a uniquing scheme. This massively reduces
memory consumption for large register allocation problems. For example, looking
at the largest interference graph in each of the SPEC2006 benchmarks (the
largest graph will always set the memory consumption high-water mark for PBQP),
the average memory reduction for the PBQP costs was 400x. That's times, not
percent. The highest was 1400x. Yikes. So - this is fixed.
"PBQP: No longer feasting upon every last byte of your RAM".
Minor details:
- Fully C++11'd. Never copy-construct another vector/matrix!
- Cute tricks with cost metadata: Metadata that is derived solely from cost
matrices/vectors is attached directly to the cost instances themselves. That way
if you unique the costs you never have to recompute the metadata. 400x less
memory means 400x less cost metadata (re)computation.
Special thanks to Arnaud de Grandmaison, who has been the source of much
encouragement, and of many very useful test cases.
This new solver forms the basis for future work, of which there's plenty to do.
I will be adding TODO notes shortly.
- Lang.
llvm-svn: 202551
After this I will set the default back to F_None. The advantage is that
before this patch forgetting to set F_Binary would corrupt a file on windows.
Forgetting to set F_Text produces one that cannot be read in notepad, which
is a better failure mode :-)
llvm-svn: 202052
This is slightly more interesting than the previous batch of changes.
Specifically:
1. We refactor getSpillWeight to take a MachineBlockFrequencyInfo (MBFI)
object. This enables us to completely encapsulate the actual manner we
use the MachineBlockFrequencyInfo to get our spill weights. This yields
cleaner code since one does not need to fetch the actual block frequency
before getting the spill weight if all one wants it the spill weight. It
also gives us access to entry frequency which we need for our
computation.
2. Instead of having getSpillWeight take a MachineBasicBlock (as one
might think) to look up the block frequency via the MBFI object, we
instead take in a MachineInstr object. The reason for this is that the
method is supposed to return the spill weight for an instruction
according to the comments around the function.
llvm-svn: 197296
Based on discussions with Lang Hames and Jakob Stoklund Olesen at the hacker's lab, and in the light of upcoming work on the PBQP register allocator, it was though that CalcSpillWeights does not need to be a pass. This change will enable to customize / tune the spill weight computation depending on the allocator.
Update the documentation style while there.
No functionnal change.
llvm-svn: 194356
The new graph structure replaces the node and edge linked lists with vectors.
Free lists (well, free vectors) are used for fast insertion/deletion.
The ultimate aim is to make PBQP graphs cheap to clone. The motivation is that
the PBQP solver destructively consumes input graphs while computing a solution,
forcing the graph to be fully reconstructed for each round of PBQP. This
imposes a high cost on large functions, which often require several rounds of
solving/spilling to find a final register allocation. If we can cheaply clone
the PBQP graph and incrementally update it between rounds then hopefully we can
reduce this cost. Further, once we begin pooling matrix/vector values (future
work), we can cache some PBQP solver metadata and share it between cloned
graphs, allowing the PBQP solver to re-use some of the computation done in
earlier rounds.
For now this is just a data structure update. The allocator and solver still
use the graph the same way as before, fully reconstructing it between each
round. I expect no material change from this update, although it may change
the iteration order of the nodes, causing ties in the solver to break in
different directions, and this could perturb the generated allocations
(hopefully in a completely benign way).
Thanks very much to Arnaud Allard de Grandmaison for encouraging me to get back
to work on this, and for a lot of discussion and many useful PBQP test cases.
llvm-svn: 194300
Based on discussions with Lang Hames and Jakob Stoklund Olesen at the hacker's lab, and in the light of upcoming work on the PBQP register allocator, it was though that CalcSpillWeights does not need to be a pass. This change will enable to customize / tune the spill weight computation depending on the allocator.
Update the documentation style while there.
No functionnal change.
llvm-svn: 194269
Track new virtual registers by register number, rather than by the live
interval created for them. This is the first step in separating the
creation of new virtual registers and new live intervals. Eventually
live intervals will be created and populated on demand after the virtual
registers have been created and used in instructions.
llvm-svn: 188434
The main advantages here are way better heuristics, taking into account not
just loop depth but also __builtin_expect and other static heuristics and will
eventually learn how to use profile info. Most of the work in this patch is
pushing the MachineBlockFrequencyInfo analysis into the right places.
This is good for a 5% speedup on zlib's deflate (x86_64), there were some very
unfortunate spilling decisions in its hottest loop in longest_match(). Other
benchmarks I tried were mostly neutral.
This changes register allocation in subtle ways, update the tests for it.
2012-02-20-MachineCPBug.ll was deleted as it's very fragile and the instruction
it looked for was gone already (but the FileCheck pattern picked up unrelated
stuff).
llvm-svn: 184105
Chandler Carruth