When deriving the range of valid values of a scalar evolution expression might
be a range [12, 8), where the upper bound is smaller than the lower bound and
where the range is expected to possibly wrap around. We theoretically could
model such a range as a union of two non-wrapping ranges, but do not do this
as of yet. Instead, we just do not derive any bounds. Before this change,
we could have obtained bounds where the maximal possible value is strictly
smaller than the minimal possible value, which is incorrect and also caused
assertions during scop modeling.
llvm-svn: 294891
To determine parameters of the matrix multiplication, we check RAW dependencies
that can be expressed using only reduction dependencies. Consequently, we
should check the reduction dependencies, if this is the case.
Reviewed-by: Tobias Grosser <tobias@grosser.es>,
Sven Verdoolaege <skimo-polly@kotnet.org>
Michael Kruse <llvm@meinersbur.de>
Differential Revision: https://reviews.llvm.org/D29814
llvm-svn: 294836
The size of the operands type is the one of the parameters required
to determine the BLIS micro-kernel. We get the size of the widest type
of the matrix multiplication operands in case there are several
different types.
Reviewed-by: Michael Kruse <llvm@meinersbur.de>
Differential Revision: https://reviews.llvm.org/D29269
llvm-svn: 294828
Instead of iterating over statements and their memory accesses to extract the
set of available base pointers, just directly iterate over all ScopArray
objects. This reflects more the actual intend of the code: collect all arrays
(and their base pointers) to emit alias information that specifies that accesses
to different arrays cannot alias.
This change removes unnecessary uses of MemoryAddress::getBaseAddr() in
preparation for https://reviews.llvm.org/D28518.
llvm-svn: 294574
There are problems with using the machine information to derive the precise
vector size on polly-amd64-linux and polly-arm-linux. We temporarily disable
the problematic run lines.
llvm-svn: 294571
Before this change we used the name of the base pointer to mark reductions. This
is imprecise as the canonical reference is the ScopArray itself and not the
basepointer of a reduction. Using the base pointer of reductions is problematic
in cases where a single ScopArray is referenced through two different base
pointers.
This change removes unnecessary uses of MemoryAddress::getBaseAddr() in
preparation for https://reviews.llvm.org/D28518.
llvm-svn: 294568
optimization
Isolate a set of partial tile prefixes to allow hoisting and sinking out of
the unrolled innermost loops produced by the optimization of the matrix
multiplication.
In case it cannot be proved that the number of loop iterations can be evenly
divided by tile sizes and we tile and unroll the point loop, the isl generates
conditional expressions. Subsequently, the conditional expressions can prevent
stores and loads of the unrolled loops from being sunk and hoisted.
The patch isolates a set of partial tile prefixes, which have exactly Mr x Nr
iterations of the two innermost loops, the result of the loop tiling performed
by the matrix multiplication optimization, where Mr and Mr are parameters of
the micro-kernel. This helps to get rid of the conditional expressions of
the unrolled innermost loops. Probably this approach can be replaced with
padding in future.
In case of, for example, the gemm from Polybench/C 3.2 and parametric loop
bounds, it helps to increase the performance from 7.98 GFlops (27.71% of
theoretical peak) to 21.47 GFlops (74.57% of theoretical peak). Hence, we
get the same performance as in case of scalar loops bounds.
It also cause compile time regression. The compile-time is increased from
0.795 seconds to 0.837 seconds in case of scalar loops bounds and from 1.222
seconds to 1.490 seconds in case of parametric loops bounds.
Reviewed-by: Michael Kruse <llvm@meinersbur.de>
Differential Revision: https://reviews.llvm.org/D29244
llvm-svn: 294564
with optimizeMatMulPattern
This patch makes ScheduleTreeOptimizer::optimizeBand return a schedule node
optimized with optimizeMatMulPattern. Otherwise, it could not use the isolate
option, because standardBandOpts could try to tile a band node with anchored
subtree and get the error, since the use of the isolate option causes any tree
containing the node to be considered anchored. Furthermore, it is not intended
to apply standard optimizations, when the matrix multiplication has been
detected.
llvm-svn: 294444
multiplication
The current identification of a SCoP statement that implement a matrix
multiplication does not help to identify different permutations of loops that
contain it and check for dependencies, which can prevent it from being
optimized. It also requires external determination of the operands of
the matrix multiplication. This patch contains the implementation of a new
algorithm that helps to avoid these issues. It also modifies the test cases
that generate matrix multiplications with linearized accesses, because
the new algorithm does not support them.
Reviewed-by: Michael Kruse <llvm@meinersbur.de>,
Tobias Grosser <tobias@grosser.es>
Differential Revision: https://reviews.llvm.org/D28357
llvm-svn: 293890
Before this change we created an additional reload in the copy of the incoming
block of a PHI node to reload the incoming value, even though the necessary
value has already been made available by the normally generated scalar loads.
In this change, we drop the code that generates this redundant reload and
instead just reuse the scalar value already available.
Besides making the generated code slightly cleaner, this change also makes sure
that scalar loads go through the normal logic, which means they can be remapped
(e.g. to array slots) and corresponding code is generated to load from the
remapped location. Without this change, the original scalar load at the
beginning of the non-affine region would have been remapped, but the redundant
scalar load would continue to load from the old PHI slot location.
It might be possible to further simplify the code in addOperandToPHI,
but this would not only mean to pull out getNewValue, but to also change the
insertion point update logic. As this did not work when trying it the first
time, this change is likely not trivial. To not introduce bugs last minute, we
postpone further simplications to a subsequent commit.
We also document the current behavior a little bit better.
Reviewed By: Meinersbur
Differential Revision: https://reviews.llvm.org/D28892
llvm-svn: 292486
We rename the test case with -metarenamer to make the variable names easier to
read and add additional check lines that verify the code we currently generate
for PHI nodes. This code is interesting as it contains a PHI node in a
non-affine sub-region, where some incoming blocks are within the non-affine
sub-region and others are outside of the non-affine subregion.
As can be seen in the check lines we currently load the PHI-node value twice.
This commit documents this behavior. In a subsequent patch we will try to
improve this.
llvm-svn: 292470
Summary:
Instead of forbidding such access functions completely, we verify that their
base pointer has been hoisted and only assert in case the base pointer was
not hoisted.
I was trying for a little while to get a test case that ensures the assert is
correctly fired in case of invariant load hoisting being disabled, but I could
not find a good way to do so, as llvm-lit immediately aborts if a command
yields a non-zero return value. As we do not generally test our asserts,
not having a test case here seems OK.
This resolves http://llvm.org/PR31494
Suggested-by: Michael Kruse <llvm@meinersbur.de>
Reviewers: efriedma, jdoerfert, Meinersbur, gareevroman, sebpop, zinob, huihuiz, pollydev
Reviewed By: Meinersbur
Differential Revision: https://reviews.llvm.org/D28798
llvm-svn: 292213
This feature is currently not supported and an explicit assert to prevent the
introduction of such accesses has been added in r282893. This test case allows
to reproduce the assert (and without the assert the miscompile) added in
r282893. It will help when adding such support at some point.
llvm-svn: 292147
If the parameters of the target cache (i.e., cache level sizes, cache level
associativities) are not specified or have wrong values, we use ones for
parameters of the macro-kernel and do not perform data-layout optimizations of
the matrix multiplication. In this patch we specify the default values of the
cache parameters to be able to apply the pattern matching optimizations even in
this case. Since there is no typical values of this parameters, we use the
parameters of Intel Core i7-3820 SandyBridge that also help to attain the
high-performance on IBM POWER System S822 and IBM Power 730 Express server.
Reviewed-by: Tobias Grosser <tobias@grosser.es>
Differential Revision: https://reviews.llvm.org/D28090
llvm-svn: 290518
Typically processor architectures do not include an L3 cache, which means that
Nc, the parameter of the micro-kernel, is, for all practical purposes,
redundant ([1]). However, its small values can cause the redundant packing of
the same elements of the matrix A, the first operand of the matrix
multiplication. At the same time, big values of the parameter Nc can cause
segmentation faults in case the available stack is exceeded.
This patch adds an option to specify the parameter Nc as a multiple of
the parameter of the micro-kernel Nr.
In case of Intel Core i7-3820 SandyBridge and the following options,
clang -O3 gemm.c -I utilities/ utilities/polybench.c -DPOLYBENCH_TIME
-march=native -mllvm -polly -mllvm -polly-pattern-matching-based-opts=true
-DPOLYBENCH_USE_SCALAR_LB -mllvm -polly-target-cache-level-associativity=8,8
-mllvm -polly-target-cache-level-sizes=32768,262144 -mllvm
-polly-target-latency-vector-fma=8
it helps to improve the performance from 11.303 GFlops/sec (39,247% of
theoretical peak) to 17.896 GFlops/sec (62,14% of theoretical peak).
Refs.:
[1] - http://www.cs.utexas.edu/users/flame/pubs/TOMS-BLIS-Analytical.pdf
Reviewed-by: Tobias Grosser <tobias@grosser.es>
Differential Revision: https://reviews.llvm.org/D28019
llvm-svn: 290256
multiplication
Previously we had two-dimensional accesses to store packed operands of
the matrix multiplication for the sake of simplicity of the packed arrays.
However, addition of the third dimension helps to simplify the corresponding
memory access, reduce the execution time of isl operations applied to it, and
consequently reduce the compile-time of Polly. For example, in case of
Intel Core i7-3820 SandyBridge and the following options,
clang -O3 gemm.c -I utilities/ utilities/polybench.c -DPOLYBENCH_TIME
-march=native -mllvm -polly -mllvm -polly-pattern-matching-based-opts=true
-DPOLYBENCH_USE_SCALAR_LB -mllvm -polly-target-cache-level-associativity=8,8
-mllvm -polly-target-cache-level-sizes=32768,262144 -mllvm
-polly-target-latency-vector-fma=7
it helps to reduce the compile-time from about 361.456 seconds to about 0.816
seconds.
Reviewed-by: Michael Kruse <llvm@meinersbur.de>,
Tobias Grosser <tobias@grosser.es>
Differential Revision: https://reviews.llvm.org/D27878
llvm-svn: 290251
To prevent copy statements from accessing arrays out of bounds, ranges of their
extension maps are restricted, according to the constraints of domains.
Reviewed-by: Michael Kruse <llvm@meinersbur.de>
Differential Revision: https://reviews.llvm.org/D25655
llvm-svn: 289815
gemm ([1]). In particular, elements of the matrix B, the second operand of
matrix multiplication, are reused between iterations of the innermost loop.
To keep the reused data in cache, only elements of matrix A, the first operand
of matrix multiplication, should be evicted during an iteration of the
innermost loop. To provide such a cache replacement policy, elements of the
matrix A can, in particular, be loaded first and, consequently, be
least-recently-used.
In our case matrices are stored in row-major order instead of column-major
order used in the BLIS implementation ([1]). One of the ways to address it is
to accordingly change the order of the loops of the loop nest. However, it
makes elements of the matrix A to be reused in the innermost loop and,
consequently, requires to load elements of the matrix B first. Since the LLVM
vectorizer always generates loads from the matrix A before loads from the
matrix B and we can not provide it. Consequently, we only change the BLIS micro
kernel and the computation of its parameters instead. In particular, reused
elements of the matrix B are successively multiplied by specific elements of
the matrix A .
Refs.:
[1] - http://www.cs.utexas.edu/users/flame/pubs/TOMS-BLIS-Analytical.pdf
Reviewed-by: Tobias Grosser <tobias@grosser.es>
Differential Revision: https://reviews.llvm.org/D25653
llvm-svn: 289806
This allows us to delinearize code such as the one below, where the array
sizes are A[][2 * n] as there are n times two elements in the innermost
dimension. Alternatively, we could try to generate another dimension for the
struct in the innermost dimension, but as the struct has constant size,
recovering this dimension is easy.
struct com {
double Real;
double Img;
};
void foo(long n, struct com A[][n]) {
for (long i = 0; i < 100; i++)
for (long j = 0; j < 1000; j++)
A[i][j].Real += A[i][j].Img;
}
int main() {
struct com A[100][1000];
foo(1000, A);
llvm-svn: 288489
Add an empty DeLICM pass, without any functional parts.
Extracting the boilerplate from the the functional part reduces the size of the
code to review (https://reviews.llvm.org/D24716)
Suggested-by: Tobias Grosser <tobias@grosser.es>
llvm-svn: 288160
We now collect:
Number of total loops
Number of loops in scops
Number of scops
Number of scops with maximal loop depth 1
Number of scops with maximal loop depth 2
Number of scops with maximal loop depth 3
Number of scops with maximal loop depth 4
Number of scops with maximal loop depth 5
Number of scops with maximal loop depth 6 and larger
Number of loops in scops (profitable scops only)
Number of scops (profitable scops only)
Number of scops with maximal loop depth 1 (profitable scops only)
Number of scops with maximal loop depth 2 (profitable scops only)
Number of scops with maximal loop depth 3 (profitable scops only)
Number of scops with maximal loop depth 4 (profitable scops only)
Number of scops with maximal loop depth 5 (profitable scops only)
Number of scops with maximal loop depth 6 and larger (profitable scops only)
These statistics are certainly completely accurate as we might drop scops
when building up their polyhedral representation, but they should give a good
indication of the number of scops we detect.
llvm-svn: 287973
Our original statistics were added before we introduced a more fine-grained
diagnostic system, but the granularity of our statistics has never been
increased accordingly. This change introduces now one statistic counter per
diagnostic to enable us to collect fine-grained statistics about who certain
scops are not detected. In case coarser grained statistics are needed, the
user is expected to combine counters manually.
llvm-svn: 287968
Introduce the new flag -polly-codegen-generate-expressions which forces Polly
to code generate AST expressions instead of using our SCEV based access
expression generation even for cases where the original memory access relation
was not changed and the SCEV based access expression could be code generated
without any issue.
This is an experimental option for better testing the isl ast expression
generation. The default behavior of Polly remains unchanged. We also exclude
a couple of cases for which the AST expression is not yet working.
llvm-svn: 287694
Drop instructions that do not influence the memory impact of a basic block.
They are not needed to reproduce the original bug (verified) and will cause
random test noise if we would decide to only model the instructions that
have visible side-effects.
llvm-svn: 287626
Add two store instructions at the end of basic blocks that are required to
reproduce the original bug to ensure we always process and model these basic
blocks. This makes this test case stable even in case we would decide to bail
out early of basic blocks which do not modify the global state. Also add
additional check lines to verify how we model the basic block.
llvm-svn: 287625
We add CHECK lines to this test case to make it easier to see the difference
between affine and non-affine memory accesses. We also change the test case to
use a parameteric index expression as otherwise our range analysis will
understand that the non-affine memory access can only access input[1],
which makes it difficult to see that the memory access is in-fact modeled as
non-affine access.
llvm-svn: 287623
Do not assume a load to be hoistable/invariant if the pointer is used by
another instruction in the SCoP that might write to memory and that is
always executed.
llvm-svn: 287272
The validity of a branch condition must be verified at the location of the
branch (the branch instruction), not the location of the icmp that is
used in the branch instruction. When verifying at the wrong location, we
may accept an icmp that is defined within a loop which itself dominates, but
does not contain the branch instruction. Such loops cannot be modeled as
we only introduce domain dimensions for surrounding loops. To address this
problem we change the scop detection to evaluate and verify SCEV expressions at
the right location.
This issue has been around since at least r179148 "scop detection: properly
instantiate SCEVs to the place where they are used", where we explicitly
set the scope to the wrong location. Before this commit the scope
was not explicitly set, which probably also resulted in the scope around the
ICmp to be choosen.
This resolves http://llvm.org/PR30989
Reported-by: Eli Friedman <efriedma@codeaurora.org>
llvm-svn: 286769
Assumptions can either be added for a given basic block, in which case the set
describing the assumptions is expected to match the dimensions of its domain.
In case no basic block is provided a parameter-only set is expected to describe
the assumption.
The piecewise expressions that are generated by the SCEVAffinator sometimes
have a zero-dimensional domain (e.g., [p] -> { [] : p <= -129 or p >= 128 }),
which looks similar to a parameter-only domain, but is still a set domain.
This change adds an assert that checks that we always pass parameter domains to
addAssumptions if BB is empty to make mismatches here fail early.
We also change visitTruncExpr to always convert to parameter sets, if BB is
null. This change resolves http://llvm.org/PR30941
Another alternative to this change would have been to inspect all code to make
sure we directly generate in the SCEV affinator parameter sets in case of empty
domains. However, this would likely complicate the code which combines parameter
and non-parameter domains when constructing a statement domain. We might still
consider doing this at some point, but as this likely requires several non-local
changes this should probably be done as a separate refactoring.
Reported-by: Eli Friedman <efriedma@codeaurora.org>
llvm-svn: 286444
Providing the context to the ast generator allows for additional simplifcations
and -- more importantly -- allows to generate loops with only partially bounded
domains, assuming the domains are bounded for all parameter configurations
that are valid as defined by the context.
This change fixes the crash reported in http://llvm.org/PR30956
The original reason why we did not include the context when generating an
AST was that CLooG and later isl used to sometimes transfer some of the
constraints that bound the size of parameters from the context into the
generated AST. This resulted in operations with very large constants, which
sometimes introduced problematic integer overflows. The latest versions of
the isl AST generator are careful to not introduce such constants.
Reported-by: Eli Friedman <efriedma@codeaurora.org>
llvm-svn: 286442
When extracting constant expressions out of SCEVs, new parameters may be
introduced, which have not been registered before. This change scans
SCEV expressions after constant extraction again to make sure newly
introduced parameters are registered.
We may for example extract the constant '8' from the expression '((8 * ((%a *
%b) + %c)) + (-8 * %a))' and obtain the expression '(((-1 + %b) * %a) + %c)'.
The new expression has a new parameter '(-1 + %b) * %a)', which was not
registered before, but must be registered to not crash.
This closes http://llvm.org/PR30953
Reported-by: Eli Friedman <efriedma@codeaurora.org>
llvm-svn: 286430
In r248701 "Allow switch instructions in SCoPs" support for switch statements
has been introduced, but support for switch statements in loop latches was
incomplete. This change completely disables switch statements in loop latches.
The original commit changed addLoopBoundsToHeaderDomain to support non-branch
terminator instructions, but this change was incorrect: it added a check for
BI != null to the if-branch of a condition, but BI was used in the else branch
es well. As a result, when a non-branch terminator instruction is encounted a
nullptr dereference is triggered. Due to missing test coverage, this bug was
overlooked.
r249273 "[FIX] Approximate non-affine loops correctly" added code to disallow
switch statements for non-affine loops, if they appear in either a loop latch
or a loop exit. We adapt this code to now prohibit switch statements in
loop latches even if the control condition is affine.
We could possibly add support for switch statements in loop latches, but such
support should be evaluated and tested separately.
This fixes llvm.org/PR30952
Reported-by: Eli Friedman <efriedma@codeaurora.org>
llvm-svn: 286426
We don't actually check whether a MemoryAccess is affine in very many
places, but one important one is in checks for aliasing.
Differential Revision: https://reviews.llvm.org/D25706
llvm-svn: 285746
When adding an llvm.memcpy instruction to AliasSetTracker, it uses the raw
source and target pointers which preserve bitcasts.
MemAccInst::getPointerOperand() also returns the raw target pointers, but
Scop::buildAliasGroups() did not for the source pointer. This lead to mismatches
between AliasSetTracker and ScopInfo on which pointer to use.
Fixed by also using raw pointers in Scop::buildAliasGroups().
llvm-svn: 285071
Integer math in LLVM IR is modular. Integer math in isl is
arbitrary-precision. Modeling LLVM IR math correctly in isl requires
either adding assumptions that math doesn't actually overflow, or
explicitly wrapping the math. However, expressions with the "nsw" flag
are special; we can pretend they're arbitrary-precision because it's
undefined behavior if the result wraps. SCEV expressions based on IR
instructions with an nsw flag also carry an nsw flag (roughly; actually,
the real rule is a bit more complicated, but the details don't matter
here).
Before this patch, SCEV flags were also overloaded with an additional
function: the ZExt code was mutating SCEV expressions as a hack to
indicate to checkForWrapping that we don't need to add assumptions to
the operand of a ZExt; it'll add explicit wrapping itself. This kind of
works... the problem is that if anything else ever touches that SCEV
expression, it'll get confused by the incorrect flags.
Instead, with this patch, we make the decision about whether to
explicitly wrap the math a bit earlier, basing the decision purely on
the SCEV expression itself, and not its users.
Differential Revision: https://reviews.llvm.org/D25287
llvm-svn: 284848
Tobias Grosser