This change is the first in a series of changes improving LLVM's Block
Frequency propogation implementation to not lose probability mass in
branchy code when propogating block frequency information from a basic
block to its successors. This patch is a simple infrastructure
improvement that does not actually modify the block frequency
algorithm. The specific changes are:
1. Changes the division algorithm used when scaling block frequencies by
branch probabilities to a short division algorithm. This gives us the
remainder for free as well as provides a nice speed boost. When I
benched the old routine and the new routine on a Sandy Bridge iMac with
disabled turbo mode performing 8192 iterations on an array of length
32768, I saw ~600% increase in speed in mean/median performance.
2. Exposes a scale method that returns a remainder. This is important so
we can ensure that when we scale a block frequency by some branch
probability BP = N/D, the remainder from the division by D can be
retrieved and propagated to other children to ensure no probability mass
is lost (more to come on this).
llvm-svn: 194950
AnalysisManager. All this method did was assert something and we have
a perfectly good way to trigger that assert from the query path.
llvm-svn: 194947
Implementing this on bigendian platforms could get strange. I added a
target hook, getStackSlotRange, per Jakob's recommendation to make
this as explicit as possible.
llvm-svn: 194942
This adds a loop rerolling pass: the opposite of (partial) loop unrolling. The
transformation aims to take loops like this:
for (int i = 0; i < 3200; i += 5) {
a[i] += alpha * b[i];
a[i + 1] += alpha * b[i + 1];
a[i + 2] += alpha * b[i + 2];
a[i + 3] += alpha * b[i + 3];
a[i + 4] += alpha * b[i + 4];
}
and turn them into this:
for (int i = 0; i < 3200; ++i) {
a[i] += alpha * b[i];
}
and loops like this:
for (int i = 0; i < 500; ++i) {
x[3*i] = foo(0);
x[3*i+1] = foo(0);
x[3*i+2] = foo(0);
}
and turn them into this:
for (int i = 0; i < 1500; ++i) {
x[i] = foo(0);
}
There are two motivations for this transformation:
1. Code-size reduction (especially relevant, obviously, when compiling for
code size).
2. Providing greater choice to the loop vectorizer (and generic unroller) to
choose the unrolling factor (and a better ability to vectorize). The loop
vectorizer can take vector lengths and register pressure into account when
choosing an unrolling factor, for example, and a pre-unrolled loop limits that
choice. This is especially problematic if the manual unrolling was optimized
for a machine different from the current target.
The current implementation is limited to single basic-block loops only. The
rerolling recognition should work regardless of how the loop iterations are
intermixed within the loop body (subject to dependency and side-effect
constraints), but the significant restriction is that the order of the
instructions in each iteration must be identical. This seems sufficient to
capture all current use cases.
This pass is not currently enabled by default at any optimization level.
llvm-svn: 194939
This patch removes most of the trivial cases of weak vtables by pinning them to
a single object file.
Differential Revision: http://llvm-reviews.chandlerc.com/D2068
Reviewed by Andy
llvm-svn: 194865
0xffff does not mean that there are 65535 sections in a COFF file but
indicates that it's a COFF import library. This patch fixes SEGV error
when an import library file is passed to llvm-readobj.
llvm-svn: 194844
Stop folding constant adds into GEP when the type size doesn't match.
Otherwise, the adds' operands are effectively being promoted, changing the
conditions of an overflow. Results are different when:
sext(a) + sext(b) != sext(a + b)
Problem originally found on x86-64, but also fixed issues with ARM and PPC,
which used similar code.
<rdar://problem/15292280>
Patch by Duncan Exon Smith!
llvm-svn: 194840
Summary:
When getConstant() is called for an expanded vector type, it is split into
multiple scalar constants which are then combined using appropriate build_vector
and bitcast operations.
In addition to the usual big/little endian differences, the case where the
element-order of the vector does not have the same endianness as the elements
themselves is also accounted for. For example, for v4i32 on big-endian MIPS,
the byte-order of the vector is <3210,7654,BA98,FEDC>. For little-endian, it is
<0123,4567,89AB,CDEF>.
Handling this case turns out to be a nop since getConstant() returns a splatted
vector (so reversing the element order doesn't change the value)
This fixes a number of cases in MIPS MSA where calling getConstant() during
operation legalization introduces illegal types (e.g. to legalize v2i64 UNDEF
into a v2i64 BUILD_VECTOR of illegal i64 zeros). It should also handle bigger
differences between illegal and legal types such as legalizing v2i64 into v8i16.
lowerMSASplatImm() in the MIPS backend no longer needs to avoid calling
getConstant() so this function has been updated in the same patch.
For the sake of transparency, the steps I've taken since the review are:
* Added 'virtual' to isVectorEltOrderLittleEndian() as requested. This revealed
that the MIPS tests were falsely passing because a polymorphic function was
not actually polymorphic in the reviewed patch.
* Fixed the tests that were now failing. This involved deleting the code to
handle the MIPS MSA element-order (which was previously doing an byte-order
swap instead of an element-order swap). This left
isVectorEltOrderLittleEndian() unused and it was deleted.
* Fixed build failures caused by rebasing beyond r194467-r194472. These build
failures involved the bset, bneg, and bclr instructions added in these commits
using lowerMSASplatImm() in a way that was no longer valid after this patch.
Some of these were fixed by calling SelectionDAG::getConstant() instead,
others were fixed by a new function getBuildVectorSplat() that provided the
removed functionality of lowerMSASplatImm() in a more sensible way.
Reviewers: bkramer
Reviewed By: bkramer
CC: llvm-commits
Differential Revision: http://llvm-reviews.chandlerc.com/D1973
llvm-svn: 194811
This is to avoid this transformation in some cases:
fold (conv (load x)) -> (load (conv*)x)
On architectures that don't natively support some vector
loads efficiently casting the load to a smaller vector of
larger types and loading is more efficient.
Patch by Micah Villmow.
llvm-svn: 194783
Including only Debug.h did not cause a compilation error, but you couldn't
do anything (like writing something with <<) to raw_ostreams returned by
llvm::dbgs() or llvm::errs() without including raw_ostream.h. So including
it from Debug.h should make sense.
Differential Revision: http://llvm-reviews.chandlerc.com/D2183
llvm-svn: 194759
This is useful for debugging issues in the BlockFrequency implementation since
one can easily visualize where probability mass and other errors occur in the
propagation.
llvm-svn: 194654
- readInt() should check all 4 bytes can be read, not just 1.
- In the event of false data in the gcno file, it was possible to index
into a non-existent index of SmallVector, causing assertion error.
llvm-svn: 194639
According to the hazy gcov documentation, it appeared to be technically
possible for lines within a block to belong to different source files.
However, upon further investigation, gcov does not actually support
multiple source files for a single block.
This change removes a level of separation between blocks and lines by
replacing the StringMap of GCOVLines with a SmallVector of ints
representing line numbers. This also means that the GCOVLines class is
no longer needed.
This paves the way for supporting the "-a" option, which will output
block information.
llvm-svn: 194637
Unified the interface for read functions. They all return a boolean
indicating if the read from file succeeded. Functions that previously
returned the read value now store it into a variable that is passed in
by reference instead. Callers will need to check the return value to
detect if an error occurred.
Also added a new test which ensures that no assertions occur when file
contains invalid data. llvm-cov should return with error code 1 upon
failure.
llvm-svn: 194635
instructions. This patch does not include the shift right and accumulate
instructions. A number of non-overloaded intrinsics have been remove in favor
of their overloaded counterparts.
llvm-svn: 194598
Accepting quotes is a property of an assembler, not of an object file. For
example, ELF can support any names for sections and symbols, but the gnu
assembler only accepts quotes in some contexts and llvm-mc in a few more.
LLVM should not produce different symbols based on a guess about which assembler
will be reading the code it is printing.
llvm-svn: 194575
This adds a new scalar pass that reads a file with samples generated
by 'perf' during runtime. The samples read from the profile are
incorporated and emmited as IR metadata reflecting that profile.
The profile file is assumed to have been generated by an external
profile source. The profile information is converted into IR metadata,
which is later used by the analysis routines to estimate block
frequencies, edge weights and other related data.
External profile information files have no fixed format, each profiler
is free to define its own. This includes both the on-disk representation
of the profile and the kind of profile information stored in the file.
A common kind of profile is based on sampling (e.g., perf), which
essentially counts how many times each line of the program has been
executed during the run.
The SampleProfileLoader pass is organized as a scalar transformation.
On startup, it reads the file given in -sample-profile-file to
determine what kind of profile it contains. This file is assumed to
contain profile information for the whole application. The profile
data in the file is read and incorporated into the internal state of
the corresponding profiler.
To facilitate testing, I've organized the profilers to support two file
formats: text and native. The native format is whatever on-disk
representation the profiler wants to support, I think this will mostly
be bitcode files, but it could be anything the profiler wants to
support. To do this, every profiler must implement the
SampleProfile::loadNative() function.
The text format is mostly meant for debugging. Records are separated by
newlines, but each profiler is free to interpret records as it sees fit.
Profilers must implement the SampleProfile::loadText() function.
Finally, the pass will call SampleProfile::emitAnnotations() for each
function in the current translation unit. This function needs to
translate the loaded profile into IR metadata, which the analyzer will
later be able to use.
This patch implements the first steps towards the above design. I've
implemented a sample-based flat profiler. The format of the profile is
fairly simplistic. Each sampled function contains a list of relative
line locations (from the start of the function) together with a count
representing how many samples were collected at that line during
execution. I generate this profile using perf and a separate converter
tool.
Currently, I have only implemented a text format for these profiles. I
am interested in initial feedback to the whole approach before I send
the other parts of the implementation for review.
This patch implements:
- The SampleProfileLoader pass.
- The base ExternalProfile class with the core interface.
- A SampleProfile sub-class using the above interface. The profiler
generates branch weight metadata on every branch instructions that
matches the profiles.
- A text loader class to assist the implementation of
SampleProfile::loadText().
- Basic unit tests for the pass.
Additionally, the patch uses profile information to compute branch
weights based on instruction samples.
This patch converts instruction samples into branch weights. It
does a fairly simplistic conversion:
Given a multi-way branch instruction, it calculates the weight of
each branch based on the maximum sample count gathered from each
target basic block.
Note that this assignment of branch weights is somewhat lossy and can be
misleading. If a basic block has more than one incoming branch, all the
incoming branches will get the same weight. In reality, it may be that
only one of them is the most heavily taken branch.
I will adjust this assignment in subsequent patches.
llvm-svn: 194566
This bug only bit the C++98 build bots because all of the actual uses
really do move. ;] But not *quite* ready to do the whole C++11 switch
yet, so clean it up. Also add a unit test that catches this immediately.
llvm-svn: 194548
more smarts in it. This is where most of the interesting logic that used
to live in the implicit-scheduling-hackery of the old pass manager will
live.
Like the previous commits, note that this is a very early prototype!
I expect substantial changes before this is ready to use.
The core of the design is the following:
- We have an AnalysisManager which can be used across a series of
passes over a module.
- The code setting up a pass pipeline registers the analyses available
with the manager.
- Individual transform passes can check than an analysis manager
provides the analyses they require in order to fail-fast.
- There is *no* implicit registration or scheduling.
- Analysis passes are different from other passes: they produce an
analysis result that is cached and made available via the analysis
manager.
- Cached results are invalidated automatically by the pass managers.
- When a transform pass requests an analysis result, either the analysis
is run to produce the result or a cached result is provided.
There are a few aspects of this design that I *know* will change in
subsequent commits:
- Currently there is no "preservation" system, that needs to be added.
- All of the analysis management should move up to the analysis library.
- The analysis management needs to support at least SCC passes. Maybe
loop passes. Living in the analysis library will facilitate this.
- Need support for analyses which are *both* module and function passes.
- Need support for pro-actively running module analyses to have cached
results within a function pass manager.
- Need a clear design for "immutable" passes.
- Need support for requesting cached results when available and not
re-running the pass even if that would be necessary.
- Need more thorough testing of all of this infrastructure.
There are other aspects that I view as open questions I'm hoping to
resolve as I iterate a bit on the infrastructure, and especially as
I start writing actual passes against this.
- Should we have separate management layers for function, module, and
SCC analyses? I think "yes", but I'm not yet ready to switch the code.
Adding SCC support will likely resolve this definitively.
- How should the 'require' functionality work? Should *that* be the only
way to request results to ensure that passes always require things?
- How should preservation work?
- Probably some other things I'm forgetting. =]
Look forward to more patches in shorter order now that this is in place.
llvm-svn: 194538
Add user-supplied C runtime and compiler-rt library functions to
llvm.compiler.used to protect them from premature optimization by
passes like -globalopt and -ipsccp. Calls to (seemingly unused)
runtime library functions can be added by -instcombine and instruction
lowering.
Patch by Duncan Exon Smith, thanks!
Fixes <rdar://problem/14740087>
llvm-svn: 194514
This reverts commit r194485.
The variable is unused in some macro instantiations, but not others. We should
probably fix clang to not warn on this.
llvm-svn: 194486
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
This is still just a skeleton. I'm trying to pull together the
experimentation I've done into committable chunks, and this is the first
coherent one. Others will follow in hopefully short order that move this
more toward a useful initial implementation. I still expect the design
to continue evolving in small ways as I work through the different
requirements and features needed here though.
Keep in mind, all of this is off by default.
Currently, this mostly exercises the use of a polymorphic smart pointer
and templates to hide the polymorphism for the pass manager from the
pass implementation. The next step will be more significant, adding the
first framework of analysis support.
llvm-svn: 194325
give the files a legacy prefix in the right directory. Use forwarding
headers in the old locations to paper over the name change for most
clients during the transitional period.
No functionality changed here! This is just clearing some space to
reduce renaming churn later on with a new system.
Even when the new stuff starts to go in, it is going to be hidden behind
a flag and off-by-default as it is still WIP and under development.
This patch is specifically designed so that very little out-of-tree code
has to change. I'm going to work as hard as I can to keep that the case.
Only direct forward declarations of the PassManager class are impacted
by this change.
llvm-svn: 194324
unique ownership smart pointer which is *deep* copyable by assuming it
can call a T::clone() method to allocate a copy of the owned data.
This is mostly useful with containers or other collections of uniquely
owned data in C++98 where they *might* copy. With C++11 we can likely
remove this in favor of move-only types and containers wrapped around
those types.
llvm-svn: 194315
This patch moves the jump address materialization inside the noop slide. This
enables patching of the materialization itself or its complete removal. This
patch also adds the ability to define scratch registers that can be used safely
by the code called from the patchpoint intrinsic. At least one scratch register
is required, because that one is used for the materialization of the jump
address. This patch depends on D2009.
Differential Revision: http://llvm-reviews.chandlerc.com/D2074
Reviewed by Andy
llvm-svn: 194306
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
The idea of the AnyReg Calling Convention is to provide the call arguments in
registers, but not to force them to be placed in a paticular order into a
specified set of registers. Instead it is up tp the register allocator to assign
any register as it sees fit. The same applies to the return value (if
applicable).
Differential Revision: http://llvm-reviews.chandlerc.com/D2009
Reviewed by Andy
llvm-svn: 194293
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
Patch by Michele Scandale!
Rewrite of the functions used to compute the backedge taken count of a
loop on LT and GT comparisons.
I decided to split the handling of LT and GT cases becasue the trick
"a > b == -a < -b" in some cases prevents the trip count computation
due to the multiplication by -1 on the two operands of the
comparison. This issue comes from the conservative computation of
value range of SCEVs: taking the negative SCEV of an expression that
have a small positive range (e.g. [0,31]), we would have a SCEV with a
fullset as value range.
Indeed, in the new rewritten function I tried to better handle the
maximum backedge taken count computation when MAX/MIN expression are
used to handle the cases where no entry guard is found.
Some test have been modified in order to check the new value correctly
(I manually check them and reasoning on possible overflow the new
values seem correct).
I finally added a new test case related to the multiplication by -1
issue on GT comparisons.
llvm-svn: 194116
One of the uses of the IsValid flag is to support default constructing
a ErrorOr that is not a Error or a Value. There is not much value in
doing that IMHO. If ErrorOr was to have a default constructor, it
should be implemented by default constructing the value, but even that
looks unnecessary.
The other use is to avoid calling destructors on moved objects. This
looks wrong. If the data being moved has non trivial treatment of
moves (an std::vector for example), it is its destructor that should
handle it, not ~ErrorOr.
With this change ErrorOr becomes a fairly simple wrapper and should
always be better than using an error_code + value in an API.
llvm-svn: 194109
This patch enables llvm-cov to correctly output the run count stored in
the GCDA file. GCOVProfiling currently does not generate this
information, so the GCDA run data had to be hacked on from a GCDA file
generated by gcc. This is corrected by a subsequent patch.
With the run and program data included, both llvm-cov and gcov produced
the same output.
llvm-svn: 194033
ErrorOr had quiet a bit of complexity and indirection to be able to hold a user
type with the error.
That feature is not used anymore. This patch removes it, it will live in svn
history if we ever need it again.
If we do need it again, IMHO there is one thing that should be done
differently: Holding extra info in the error is not a property a function also
returning a value or not. The ability to hold extra info should be in the error
type and ErrorOr templated over it so that we don't need the funny looking
ErrorOr<void>.
llvm-svn: 194030
As with the other loop unrolling parameters (the unrolling threshold, partial
unrolling, etc.) runtime unrolling can now also be controlled via the
constructor. This will be necessary for moving non-trivial unrolling late in
the pass manager (after loop vectorization).
No functionality change intended.
llvm-svn: 194027
stack traces by default if you use PrettyStackTraceProgram, so that existing LLVM-based
tools will continue to get it without any changes.
llvm-svn: 193971
Michael Gottesman