This is essentially a ground up re-think of the SROA pass in LLVM. It
was initially inspired by a few problems with the existing pass:
- It is subject to the bane of my existence in optimizations: arbitrary
thresholds.
- It is overly conservative about which constructs can be split and
promoted.
- The vector value replacement aspect is separated from the splitting
logic, missing many opportunities where splitting and vector value
formation can work together.
- The splitting is entirely based around the underlying type of the
alloca, despite this type often having little to do with the reality
of how that memory is used. This is especially prevelant with unions
and base classes where we tail-pack derived members.
- When splitting fails (often due to the thresholds), the vector value
replacement (again because it is separate) can kick in for
preposterous cases where we simply should have split the value. This
results in forming i1024 and i2048 integer "bit vectors" that
tremendously slow down subsequnet IR optimizations (due to large
APInts) and impede the backend's lowering.
The new design takes an approach that fundamentally is not susceptible
to many of these problems. It is the result of a discusison between
myself and Duncan Sands over IRC about how to premptively avoid these
types of problems and how to do SROA in a more principled way. Since
then, it has evolved and grown, but this remains an important aspect: it
fixes real world problems with the SROA process today.
First, the transform of SROA actually has little to do with replacement.
It has more to do with splitting. The goal is to take an aggregate
alloca and form a composition of scalar allocas which can replace it and
will be most suitable to the eventual replacement by scalar SSA values.
The actual replacement is performed by mem2reg (and in the future
SSAUpdater).
The splitting is divided into four phases. The first phase is an
analysis of the uses of the alloca. This phase recursively walks uses,
building up a dense datastructure representing the ranges of the
alloca's memory actually used and checking for uses which inhibit any
aspects of the transform such as the escape of a pointer.
Once we have a mapping of the ranges of the alloca used by individual
operations, we compute a partitioning of the used ranges. Some uses are
inherently splittable (such as memcpy and memset), while scalar uses are
not splittable. The goal is to build a partitioning that has the minimum
number of splits while placing each unsplittable use in its own
partition. Overlapping unsplittable uses belong to the same partition.
This is the target split of the aggregate alloca, and it maximizes the
number of scalar accesses which become accesses to their own alloca and
candidates for promotion.
Third, we re-walk the uses of the alloca and assign each specific memory
access to all the partitions touched so that we have dense use-lists for
each partition.
Finally, we build a new, smaller alloca for each partition and rewrite
each use of that partition to use the new alloca. During this phase the
pass will also work very hard to transform uses of an alloca into a form
suitable for promotion, including forming vector operations, speculating
loads throguh PHI nodes and selects, etc.
After splitting is complete, each newly refined alloca that is
a candidate for promotion to a scalar SSA value is run through mem2reg.
There are lots of reasonably detailed comments in the source code about
the design and algorithms, and I'm going to be trying to improve them in
subsequent commits to ensure this is well documented, as the new pass is
in many ways more complex than the old one.
Some of this is still a WIP, but the current state is reasonbly stable.
It has passed bootstrap, the nightly test suite, and Duncan has run it
successfully through the ACATS and DragonEgg test suites. That said, it
remains behind a default-off flag until the last few pieces are in
place, and full testing can be done.
Specific areas I'm looking at next:
- Improved comments and some code cleanup from reviews.
- SSAUpdater and enabling this pass inside the CGSCC pass manager.
- Some datastructure tuning and compile-time measurements.
- More aggressive FCA splitting and vector formation.
Many thanks to Duncan Sands for the thorough final review, as well as
Benjamin Kramer for lots of review during the process of writing this
pass, and Daniel Berlin for reviewing the data structures and algorithms
and general theory of the pass. Also, several other people on IRC, over
lunch tables, etc for lots of feedback and advice.
llvm-svn: 163883
This patch implements ProfileDataLoader which loads profile data generated by
-insert-edge-profiling and updates branch weight metadata accordingly.
Patch by Alastair Murray.
llvm-svn: 162799
This is still a work in progress.
Out-of-order CPUs usually execute instructions from multiple basic
blocks simultaneously, so it is necessary to look at longer traces when
estimating the performance effects of code transformations.
The MachineTraceMetrics analysis will pick a typical trace through a
given basic block and provide performance metrics for the trace. Metrics
will include:
- Instruction count through the trace.
- Issue count per functional unit.
- Critical path length, and per-instruction 'slack'.
These metrics can be used to determine the performance limiting factor
when executing the trace, and how it will be affected by a code
transformation.
Initially, this will be used by the early if-conversion pass.
llvm-svn: 160796
This pass performs if-conversion on SSA form machine code by
speculatively executing both sides of the branch and using a cmov
instruction to select the result. This can help lower the number of
branch mispredictions on architectures like x86 that don't have
predicable instructions.
The current implementation is very aggressive, and causes regressions on
mosts tests. It needs good heuristics that have yet to be implemented.
llvm-svn: 159694
The LiveRegMatrix represents the live range of assigned virtual
registers in a Live interval union per register unit. This is not
fundamentally different from the interference tracking in RegAllocBase
that both RABasic and RAGreedy use.
The important differences are:
- LiveRegMatrix tracks interference per register unit instead of per
physical register. This makes interference checks cheaper and
assignments slightly more expensive. For example, the ARM D7 reigster
has 24 aliases, so we would check 24 physregs before assigning to one.
With unit-based interference, we check 2 units before assigning to 2
units.
- LiveRegMatrix caches regmask interference checks. That is currently
duplicated functionality in RABasic and RAGreedy.
- LiveRegMatrix is a pass which makes it possible to insert
target-dependent passes between register allocation and rewriting.
Such passes could tweak the register assignments with interference
checking support from LiveRegMatrix.
Eventually, RABasic and RAGreedy will be switched to LiveRegMatrix.
llvm-svn: 158255
OK, not really. We don't want to reintroduce the old rewriter hacks.
This patch extracts virtual register rewriting as a separate pass that
runs after the register allocator. This is possible now that
CodeGen/Passes.cpp can configure the full optimizing register allocator
pipeline.
The rewriter pass uses register assignments in VirtRegMap to rewrite
virtual registers to physical registers, and it inserts kill flags based
on live intervals.
These finalization steps are the same for the optimizing register
allocators: RABasic, RAGreedy, and PBQP.
llvm-svn: 158244
Besides adding the new insertPass function, this patch uses it to
enhance the existing -print-machineinstrs so that the MachineInstrs
after a specific pass can be printed.
Patch by Bin Zeng!
llvm-svn: 157655
Moving toward a uniform style of pass definition to allow easier target configuration.
Globally declare Pass ID.
Globally declare pass initializer.
Use INITIALIZE_PASS consistently.
Add a call to the initializer from CodeGen.cpp.
Remove redundant "createPass" functions and "getPassName" methods.
While cleaning up declarations, cleaned up comments (sorry for large diff).
llvm-svn: 150100
This is the initial checkin of the basic-block autovectorization pass along with some supporting vectorization infrastructure.
Special thanks to everyone who helped review this code over the last several months (especially Tobias Grosser).
llvm-svn: 149468
opportunities that only present themselves after late optimizations
such as tail duplication .e.g.
## BB#1:
movl %eax, %ecx
movl %ecx, %eax
ret
The register allocator also leaves some of them around (due to false
dep between copies from phi-elimination, etc.)
This required some changes in codegen passes. Post-ra scheduler and the
pseudo-instruction expansion passes have been moved after branch folding
and tail merging. They were before branch folding before because it did
not always update block livein's. That's fixed now. The pass change makes
independently since we want to properly schedule instructions after
branch folding / tail duplication.
rdar://10428165
rdar://10640363
llvm-svn: 147716
to finalize MI bundles (i.e. add BUNDLE instruction and computing register def
and use lists of the BUNDLE instruction) and a pass to unpack bundles.
- Teach more of MachineBasic and MachineInstr methods to be bundle aware.
- Switch Thumb2 IT block to MI bundles and delete the hazard recognizer hack to
prevent IT blocks from being broken apart.
llvm-svn: 146542
the mailing list. Suggestions for other statistics to collect would be
awesome. =]
Currently these are implemented as a separate pass guarded by a separate
flag. I'm not thrilled by that, but I wanted to be able to collect the
statistics for the old code placement as well as the new in order to
have a point of comparison. I'm planning on folding them into the single
pass if / when there is only one pass of interest.
llvm-svn: 143537
block frequency analyses. This differs substantially from the existing
block-placement pass in LLVM:
1) It operates on the Machine-IR in the CodeGen layer. This exposes much
more (and more precise) information and opportunities. Also, the
results are more stable due to fewer transforms ocurring after the
pass runs.
2) It uses the generalized probability and frequency analyses. These can
model static heuristics, code annotation derived heuristics as well
as eventual profile loading. By basing the optimization on the
analysis interface it can work from any (or a combination) of these
inputs.
3) It uses a more aggressive algorithm, both building chains from tho
bottom up to maximize benefit, and using an SCC-based walk to layout
chains of blocks in a profitable ordering without O(N^2) iterations
which the old pass involves.
The pass is currently gated behind a flag, and not enabled by default
because it still needs to grow some important features. Most notably, it
needs to support loop aligning and careful layout of loop structures
much as done by hand currently in CodePlacementOpt. Once it supports
these, and has sufficient testing and quality tuning, it should replace
both of these passes.
Thanks to Nick Lewycky and Richard Smith for help authoring & debugging
this, and to Jakob, Andy, Eric, Jim, and probably a few others I'm
forgetting for reviewing and answering all my questions. Writing
a backend pass is *sooo* much better now than it used to be. =D
llvm-svn: 142641
patch brings numerous advantages to LLVM. One way to look at it
is through diffstat:
109 files changed, 3005 insertions(+), 5906 deletions(-)
Removing almost 3K lines of code is a good thing. Other advantages
include:
1. Value::getType() is a simple load that can be CSE'd, not a mutating
union-find operation.
2. Types a uniqued and never move once created, defining away PATypeHolder.
3. Structs can be "named" now, and their name is part of the identity that
uniques them. This means that the compiler doesn't merge them structurally
which makes the IR much less confusing.
4. Now that there is no way to get a cycle in a type graph without a named
struct type, "upreferences" go away.
5. Type refinement is completely gone, which should make LTO much MUCH faster
in some common cases with C++ code.
6. Types are now generally immutable, so we can use "Type *" instead
"const Type *" everywhere.
Downsides of this patch are that it removes some functions from the C API,
so people using those will have to upgrade to (not yet added) new API.
"LLVM 3.0" is the right time to do this.
There are still some cleanups pending after this, this patch is large enough
as-is.
llvm-svn: 134829
BranchProbabilityInfo (expect setEdgeWeight which is not available here).
Branch Weights are kept in MachineBasicBlocks. To turn off this analysis
set -use-mbpi=false.
llvm-svn: 133184
BranchProbabilityInfo provides an interface for IR passes to query the
likelihood that control follows a CFG edge. This patch provides an
initial implementation of static branch predication that will populate
BranchProbabilityInfo for branches with no external profile
information using very simple heuristics. It currently isn't hooked up
to any external profile data, so static prediction does all the work.
llvm-svn: 132613
Use debug info in the IR to find the directory/file:line:col. Each time that location changes, bump a counter.
Unlike the existing profiling system, we don't try to look at argv[], and thusly don't require main() to be present in the IR. This matches GCC's technique where you specify the profiling flag when producing each .o file.
The runtime library is minimal, currently just calling printf at program shutdown time. The API is designed to make it possible to emit GCOV data later on.
llvm-svn: 129340
query about available library functions. For now this just has
memset_pattern16, which exists on darwin, but it can be extended for a
bunch of other things in the future.
llvm-svn: 125965
Modified patch by Adam Preuss.
This builds on the existing framework for block tracing, edge profiling and optimal edge profiling.
See -help-hidden for new flags.
For documentation, see the technical report "Implementation of Path Profiling..." in llvm.org/pubs.
llvm-svn: 124515
