Add the ability to tag certain memory accesses as those belonging to
Fortran arrays. We do this by pattern matching against known patterns
of Dragonegg's LLVM IR output from Fortran code.
Fortran arrays have metadata stored with them in a struct. This struct
is called the "Fortran array descriptor", and a reference to this is
stored in each MemoryAccess.
Differential Revision: https://reviews.llvm.org/D32639
llvm-svn: 302653
Summary:
In case two arrays share base pointers in the same invariant load equivalence
class, we canonicalize all memory accesses to the first of these arrays
(according to their order in the equivalence class).
This enables us to optimize kernels such as boost::ublas by ensuring that
different references to the C array are interpreted as accesses to the same
array. Before this change the runtime alias check for ublas would fail, as it
would assume models of the C array with differing (but identically valued) base
pointers would reference distinct regions of memory whereas the referenced
memory regions were indeed identical.
As part of this change we remove most of the MemoryAccess::get*BaseAddr
interface. We removed already all references to get*BaseAddr in previous
commits to ensure that no code relies on matching base pointers between
memory accesses and scop arrays -- except for three remaining uses where we
need the original base pointer. We document for these situations that
MemoryAccess::getOriginalBaseAddr may return a base pointer that is distinct
to the base pointer of the scop array referenced by this memory access.
Reviewers: sebpop, Meinersbur, zinob, gareevroman, pollydev, huihuiz, efriedma, jdoerfert
Reviewed By: Meinersbur
Subscribers: etherzhhb
Tags: #polly
Differential Revision: https://reviews.llvm.org/D28518
llvm-svn: 302636
Summary: PPCGCodeGeneration now attaches the size of the kernel launch parameters at the end of the parameter list. For the existing CUDA Runtime, this gets ignored, but the OpenCL Runtime knows to check for kernel-argument size at the end of the parameter list. (The resulting parameters list is twice as long. This has been accounted for in the corresponding test cases).
Reviewers: grosser, Meinersbur, bollu
Reviewed By: bollu
Subscribers: nemanjai, yaxunl, Anastasia, pollydev, llvm-commits
Tags: #polly
Differential Revision: https://reviews.llvm.org/D32961
llvm-svn: 302515
Summary:
When compiling for GPU, one can now choose to compile for OpenCL or CUDA,
with the corresponding polly-gpu-runtime flag (libopencl / libcudart). The
GPURuntime library (GPUJIT) has been extended with the OpenCL Runtime library
for that purpose, correctly choosing the corresponding library calls to the
option chosen when compiling (via different initialization calls).
Additionally, a specific GPU Target architecture can now be chosen with -polly-gpu-arch (only nvptx64 implemented thus far).
Reviewers: grosser, bollu, Meinersbur, etherzhhb, singam-sanjay
Reviewed By: grosser, Meinersbur
Subscribers: singam-sanjay, llvm-commits, pollydev, nemanjai, mgorny, yaxunl, Anastasia
Tags: #polly
Differential Revision: https://reviews.llvm.org/D32431
llvm-svn: 302379
Extend the Knowledge class to store information about the contents
of array elements and which values are written. Two knowledges do
not conflict the known content is the same. The content information
if computed from writes to and loads from the array elements, and
represented by "ValInst": isl spaces that compare equal if the value
represented is the same.
Differential Revision: https://reviews.llvm.org/D31247
llvm-svn: 302339
SCoPs with unfeasible runtime context are thrown away and therefore
do not need their uses verified.
The added test case requires a complexity limit to exceed.
Normally, error statements are removed from the SCoP and for that
reason are skipped during the verification. If there is a unfeasible
runtime context (here: because of the complexity limit being reached),
the removal of error statements and other SCoP construction steps are
skipped to not waste time. Error statements are not modeled in SCoPs
and therefore have no requirements on whether the scalars used in
them are available.
llvm-svn: 302234
Since r294891, in MemoryAccess::computeBoundsOnAccessRelation(), we skip
manually bounding the access relation in case the parameter of the load
instruction is already a wrapped set. Later on we assume that the lower
bound on the set is always smaller or equal to the upper bound on the
set. Bug 32715 manages to construct a sign wrapped set, in which case
the assertion does not necessarily hold. Fix this by handling a sign
wrapped set similar to a normal wrapped set, that is skipping the
computation.
Contributed-by: Maximilian Falkenstein <falkensm@student.ethz.ch>
Reviewers: grosser
Subscribers: pollydev, llvm-commits
Tags: #Polly
Differential Revision: https://reviews.llvm.org/D32893
llvm-svn: 302231
This reverts commit 17a84e414adb51ee375d14836d4c2a817b191933.
Patches should have been submitted in the order of:
1. D32852
2. D32854
3. D32431
I mistakenly pushed D32431(3) first. Reverting to push in the correct
order.
llvm-svn: 302217
Summary:
When compiling for GPU, one can now choose to compile for OpenCL or CUDA,
with the corresponding polly-gpu-runtime flag (libopencl / libcudart). The
GPURuntime library (GPUJIT) has been extended with the OpenCL Runtime library
for that purpose, correctly choosing the corresponding library calls to the
option chosen when compiling (via different initialization calls).
Additionally, a specific GPU Target architecture can now be chosen with -polly-gpu-arch (only nvptx64 implemented thus far).
Reviewers: grosser, bollu, Meinersbur, etherzhhb, singam-sanjay
Reviewed By: grosser, Meinersbur
Subscribers: singam-sanjay, llvm-commits, pollydev, nemanjai, mgorny, yaxunl, Anastasia
Tags: #polly
Differential Revision: https://reviews.llvm.org/D32431
llvm-svn: 302215
The test subdirectory POLLY_TEST_DIRECTORIES was heavily outdated and
only used in out-of-LLVM-tree builds
(to generate polly-test-${subdir} targets).
llvm-svn: 302142
This makes sure we still test the case that a PHI-NODE cannot be analyzed by
scalar evolution and consequently must be code generated explicitly. As
Michael's optimization triggers only on a very specific "add %iv, %step"
pattern, just changing 'add' to 'mul' adds back test coverage.
llvm-svn: 302132
LLVM-IR names are commonly available in debug builds, but often not in release
builds. Hence, using LLVM-IR names to identify statements or memory reference
results makes the behavior of Polly depend on the compile mode. This is
undesirable. Hence, we now just number the statements instead of using LLVM-IR
names to identify them (this issue has previously been brought up by Zino
Benaissa).
However, as LLVM-IR names help in making test cases more readable, we add an
option '-polly-use-llvm-names' to still use LLVM-IR names. This flag is by
default set in the polly tests to make test cases more readable.
This change reduces the time in ScopInfo from 32 seconds to 2 seconds for the
following test case provided by Eli Friedman <efriedma@codeaurora.org> (already
used in one of the previous commits):
struct X { int x; };
void a();
#define SIG (int x, X **y, X **z)
typedef void (*fn)SIG;
#define FN { for (int i = 0; i < x; ++i) { (*y)[i].x += (*z)[i].x; } a(); }
#define FN5 FN FN FN FN FN
#define FN25 FN5 FN5 FN5 FN5
#define FN125 FN25 FN25 FN25 FN25 FN25
#define FN250 FN125 FN125
#define FN1250 FN250 FN250 FN250 FN250 FN250
void x SIG { FN1250 }
For a larger benchmark I have on-hand (10000 loops), this reduces the time for
running -polly-scops from 5 minutes to 4 minutes, a reduction by 20%.
The reason for this large speedup is that our previous use of printAsOperand
had a quadratic cost, as for each printed and unnamed operand the full function
was scanned to find the instruction number that identifies the operand.
We do not need to adjust the way memory reference ids are constructured, as
they do not use LLVM values.
Reviewed by: efriedma
Tags: #polly
Differential Revision: https://reviews.llvm.org/D32789
llvm-svn: 302072
- Fixes breakage from commit 5536f.
- Interference with commit 764f3 caused testcase to fail. Reverting
764f3 allows commit 5536f to succeed.
- Generated kernel code was slightly different due to 764f3, which
caused testcase to fail.
llvm-svn: 302021
Before this change a memory reference identifier had the form:
<STMT>_<ACCESSTYPE><ID>_<MEMREF>, e.g., Stmt_bb9_Write0_MemRef_tmp11
After this change, we use the format:
<STMT>_<ACCESSTYPE><ID>, e.g., Stmt_bb9_Write0
The name of the array that is accessed through a memory reference is not
necessary to uniquely identify a memory reference, but was only added to
provide additional information for debugging. We drop this information now
for the following two reasons:
1) This shortens the names and consequently improves readability
2) This removes a second location where we decide on the name of a scop array,
leaving us only with the location where the actual scop array is created.
Having after 2) only a single location to name scop arrays will allow us to
change the naming convention of scop arrays more easily, which we will do
in a future commit to reduce compilation time.
llvm-svn: 302004
This makes it easier to read and possibly even modify the test cases, as there
is no need to keep the variable increment in steps of one. More importantly, by
using explicit variable names we do not need to rely on the implicit numbering
of statements when dumping the scop information.
This makes it easier to read and possibly even modify the test cases.
Furthermore, by using explicit variables we do not need to rely on the implicit
numbering of statements when dumping the scop information. In a future commit,
this implicit numbering will likely not be used any more to refer to LLVM-IR
values as it is very expensive to construct.
llvm-svn: 301689
generation.
This needs changes to GPURuntime to expose synchronization between host
and device.
1. Needs better function naming, I want a better name than
"getOrCreateManagedDeviceArray"
2. DeviceAllocations is used by both the managed memory and the
non-managed memory path. This exploits the fact that the two code paths
are never run together. I'm not sure if this is the best design decision
Reviewed by: PhilippSchaad
Tags: #polly
Differential Revision: https://reviews.llvm.org/D32215
llvm-svn: 301640
When we introduced in r297375 support for hoisting loads that are known
to be dereferencable without any conditional guard, we forgot to keep the check
to verify that no other write into the very same location exists. This
change ensures now that dereferencable loads are allowed to access everything,
but can only be hoisted in case no conflicting write exists.
This resolves llvm.org/PR32778
Reported-by: Huihui Zhang <huihuiz@codeaurora.org>
llvm-svn: 301582
Added a small change to the way pointer arguments are set in the kernel
code generation. The way the pointer is retrieved now, specifically requests
global address space to be annotated. This is necessary, if the IR should be
run through NVPTX to generate OpenCL compatible PTX.
The changes do not affect the PTX Strings generated for the CUDA target
(nvptx64-nvidia-cuda), but are necessary for OpenCL (nvptx64-nvidia-nvcl).
Additionally, the data layout has been updated to what the NVPTX Backend requests/recommends.
Contributed-by: Philipp Schaad
Reviewers: Meinersbur, grosser, bollu
Reviewed By: grosser, bollu
Subscribers: jlebar, pollydev, llvm-commits, nemanjai, yaxunl, Anastasia
Tags: #polly
Differential Revision: https://reviews.llvm.org/D32215
llvm-svn: 301299
Earlier, the call to buildFlow was:
WAR = buildFlow(Write, Read, MustWrite, Schedule).
This meant that Read could block another Read, since must-sources can
block each other.
Fixed the call to buildFlow to correctly compute Read. The resulting
code needs to do some ISL juggling to get the output we want.
Bug report: https://bugs.llvm.org/show_bug.cgi?id=32623
Reviewers: Meinersbur
Tags: #polly
Differential Revision: https://reviews.llvm.org/D32011
llvm-svn: 301266
The isl unittest modified its PATH variable to point to the LLVM bin dir.
When building out-of-LLVM-tree, it does not contain the
polly-isl-test executable, hence the test fails.
Ensure that the polly-isl-test is written to a bin directory in the
build root, just like it would happen in an inside-LLVM build.
Then, change PATH to include that dir such that the executable in it
is prioritized before any other location.
llvm-svn: 301096
Dimensions of band nodes can be implicitly permuted by the algorithm applied
during the schedule generation.
For example, in case of the following matrix-matrix multiplication,
for (i = 0; i < 1024; i++)
for (k = 0; k < 1024; k++)
for (j = 0; j < 1024; j++)
C[i][j] += A[i][k] * B[k][j];
it can produce the following schedule tree
domain: "{ Stmt_for_body6[i0, i1, i2] : 0 <= i0 <= 1023 and 0 <= i1 <= 1023 and
0 <= i2 <= 1023 }"
child:
schedule: "[{ Stmt_for_body6[i0, i1, i2] -> [(i0)] },
{ Stmt_for_body6[i0, i1, i2] -> [(i1)] },
{ Stmt_for_body6[i0, i1, i2] -> [(i2)] }]"
permutable: 1
coincident: [ 1, 1, 0 ]
The current implementation of the pattern matching optimizations relies on the
initial ordering of dimensions. Otherwise, it can produce the miscompilation
(e.g., [1]).
This patch helps to restore the initial ordering of dimensions by recreating
the band node when the corresponding conditions are satisfied.
Refs.:
[1] - https://bugs.llvm.org/show_bug.cgi?id=32500
Reviewed-by: Michael Kruse <llvm@meinersbur.de>
Differential Revision: https://reviews.llvm.org/D31741
llvm-svn: 299662
Because Polly exposes parameters that directly influence tile size
calculations, one can setup situations like divide-by-zero.
Check against a possible divide-by-zero in getMacroKernelParams
and return early.
Also assert at the end of getMacroKernelParams that the block sizes
computed for matrices are positive (>= 1).
Tags: #polly
Differential Revision: https://reviews.llvm.org/D31708
llvm-svn: 299633
The current StackColoring algorithm does not correctly handle the
situation when some, but not all paths from a BB to the entry node
cross a llvm.lifetime.start. According to an interpretation of the
language reference at
http://llvm.org/docs/LangRef.html#llvm-lifetime-start-intrinsic
this might be correct, but it would cost too much effort to handle
in StackColoring.
To be on the safe side, remove all lifetime markers even in the original
code version (they have never been copied to the optimized version)
to ensure that no path to the entry block will cross a
llvm.lifetime.start.
The same principle applies to paths the a function return and the
llvm.lifetime.end marker, so we remove them as well.
This fixes llvm.org/PR32251.
Also see the discussion at
http://lists.llvm.org/pipermail/llvm-dev/2017-March/111551.html
llvm-svn: 299585
= Change of WAR, WAW generation: =
- `buildFlow(Sink, MustSource, MaySource, Sink)` treates any flow of the form
`sink <- may source <- must source` as a *may* dependence.
- we used to call:
```lang=cpp, name=old-flow-call.cpp
Flow = buildFlow(MustWrite, MustWrite, Read, Schedule);
WAW = isl_union_flow_get_must_dependence(Flow);
WAR = isl_union_flow_get_may_dependence(Flow);
```
- This caused some WAW dependences to be treated as WAR dependences.
- Incorrect semantics.
- Now, we call WAR and WAW correctly.
== Correct WAW: ==
```lang=cpp, name=new-waw-call.cpp
Flow = buildFlow(Write, MustWrite, MayWrite, Schedule);
WAW = isl_union_flow_get_may_dependence(Flow);
isl_union_flow_free(Flow);
```
== Correct WAR: ==
```lang=cpp, name=new-war-call.cpp
Flow = buildFlow(Write, Read, MustaWrite, Schedule);
WAR = isl_union_flow_get_must_dependence(Flow);
isl_union_flow_free(Flow);
```
- We want the "shortest" WAR possible (exact dependences).
- We mark all the *must-writes* as may-source, reads as must-souce.
- Then, we ask for *must* dependence.
- This removes all the reads that flow through a *must-write*
before reaching a sink.
- Note that we only block ealier writes with *must-writes*. This is
intuitively correct, as we do not want may-writes to block
must-writes.
- Leaves us with direct (R -> W).
- This affects reduction generation since RED is built using WAW and WAR.
= New StrictWAW for Reductions: =
- We used to call:
```lang=cpp,name=old-waw-war-call.cpp
Flow = buildFlow(MustWrite, MustWrite, Read, Schedule);
WAW = isl_union_flow_get_must_dependence(Flow);
WAR = isl_union_flow_get_may_dependence(Flow);
```
- This *is* the right model of WAW we need for reductions, just not in general.
- Reductions need to track only *strict* WAW, without any interfering reductions.
= Explanation: Why the new WAR dependences in tests are correct: =
- We no longer set WAR = WAR - WAW
- Hence, we will have WAR dependences that were originally removed.
- These may look incorrect, but in fact make sense.
== Code: ==
```lang=llvm, name=new-war-dependence.ll
; void manyreductions(long *A) {
; for (long i = 0; i < 1024; i++)
; for (long j = 0; j < 1024; j++)
; S0: *A += 42;
;
; for (long i = 0; i < 1024; i++)
; for (long j = 0; j < 1024; j++)
; S1: *A += 42;
;
```
=== WAR dependence: ===
{ S0[1023, 1023] -> S1[0, 0] }
- Between `S0[1023, 1023]` and `S1[0, 0]`, we will have the dependences:
```lang=cpp, name=dependence-incorrect, counterexample
S0[1023, 1023]:
*-- tmp = *A (load0)--*
WAR 2 add = tmp + 42 |
*-> *A = add (store0) |
WAR 1
S1[0, 0]: |
tmp = *A (load1) |
add = tmp + 42 |
A = add (store1)<-*
```
- One may assume that WAR2 *hides* WAR1 (since store0 happens before
store1). However, within a statement, Polly has no idea about the
ordering of loads and stores.
- Hence, according to Polly, the code may have looked like this:
```lang=cpp, name=dependence-correct
S0[1023, 1023]:
A = add (store0)
tmp = A (load0) ---*
add = A + 42 |
WAR 1
S1[0, 0]: |
tmp = A (load1) |
add = A + 42 |
A = add (store1) <-*
```
- So, Polly generates (correct) WAR dependences. It does not make sense
to remove these dependences, since they are correct with respect to
Polly's model.
Reviewers: grosser, Meinersbur
tags: #polly
Differential revision: https://reviews.llvm.org/D31386
llvm-svn: 299429
Add support for -polly-codegen-perf-monitoring. When performance monitoring
is enabled, we emit performance monitoring code during code generation that
prints after program exit statistics about the total number of cycles executed
as well as the number of cycles spent in scops. This gives an estimate on how
useful polyhedral optimizations might be for a given program.
Example output:
Polly runtime information
-------------------------
Total: 783110081637
Scops: 663718949365
In the future, we might also add functionality to measure how much time is spent
in optimized scops and how many cycles are spent in the fallback code.
Reviewers: bollu,sebpop
Tags: #polly
Differential Revision: https://reviews.llvm.org/D31599
llvm-svn: 299359
Trivial fix for two testcases. When Polly isn't linked into opt,
independent of whether it's built in-tree or not, these testcases forget
to load the appropriate library.
Contributed-by: Philip Pfaffe <philip.pfaffe@gmail.com>
Differential Revision: https://reviews.llvm.org/D31596
llvm-svn: 299357
Introduce another level of alias metadata to distinguish the individual
non-aliasing accesses that have inter iteration alias-free base pointers
marked with "Inter iteration alias-free" mark nodes. It can be used to,
for example, distinguish different stores (loads) produced by unrolling of
the innermost loops and, subsequently, sink (hoist) them by LICM.
Reviewed-by: Tobias Grosser <tobias@grosser.es>
Differential Revision: https://reviews.llvm.org/D30606
llvm-svn: 298510
Map the new load to the base pointer of the invariant load hoisted load
to be able to find the alias information for it.
Reviewed-by: Tobias Grosser <tobias@grosser.es>
Differential Revision: https://reviews.llvm.org/D30605
llvm-svn: 298507
When not adding constraints on parameters using -polly-ignore-parameter-bounds,
the context may not necessarily list all parameter dimensions. To support code
generation in this situation, we now always iterate over the actual parameter
list, rather than relying on the context to list all parameter dimensions.
llvm-svn: 298197
In commit r219005 lifetime markers have been introduced to mark the lifetime of
the OpenMP context data structure. However, their use seems incorrect and
recently caused a miscompile in ASC_Sequoia/CrystalMk after r298053 which was
not at all related to r298053. r298053 only caused a change in the loop order,
as this change resulted in a different isl internal representation which caused
the scheduler to derive a different schedule. This change then caused the IR to
change, which apparently created a pattern in which LLVM exploites the lifetime
markers. It seems we are using the OpenMP context outside of the lifetime
markers. Even though CrystalMk could probably be fixed by expanding the scope of
the lifetime markers, it is not clear what happens in case the OpenMP function
call is in a loop which will cause a sequence of starting and ending lifetimes.
As it is unlikely that the lifetime markers give any performance benefit, we
just drop them to remove complexity.
llvm-svn: 298192
ScopInfo's normal profitability heuristic considers SCoPs where all
statements have scalar writes as not profitably optimizable and
invalidate the SCoP in that case. However, -polly-delicm and
-polly-simplify may be able to remove some of the scalar writes such
that the flag -polly-unprofitable-scalar-accs=false allows disabling
that part of the heuristic.
In cases where DeLICM (or other passes after ScopInfo) are not
successful in removing scalar writes, the SCoP is still not profitably
optimizable. The schedule optimizer would again try computing another
schedule, resulting in slower compilation.
The -polly-prune-unprofitable pass applies the profitability heuristic
again before the schedule optimizer Polly can still bail out even with
-polly-unprofitable-scalar-accs=false.
Differential Revision: https://reviews.llvm.org/D31033
llvm-svn: 298080
Dependences::calculateDependences.
This ensures that we handle may-writes correctly when building
dependence information. Also add a test case checking correctness of
may-write information. Not handling it before was an oversight.
Differential Revision: https://reviews.llvm.org/D31075
llvm-svn: 298074
In ScheduleOptimizer::isTileableBand(), allow the case in which
the band node's child is an isl_schedule_sequence_node and its
grandchildren isl_schedule_leaf_nodes. This case can arise when
two or more statements are fused by the isl scheduler.
The tile_after_fusion.ll test has two statements in separate
loop nests and checks whether they are tiled after being fused
when polly-opt-fusion equals "max".
Reviewers: grosser
Subscribers: gareevroman, pollydev
Tags: #polly
Contributed-by: Theodoros Theodoridis <theodort@student.ethz.ch>
Differential Revision: https://reviews.llvm.org/D30815
llvm-svn: 297587
This new pass removes unnecessary accesses and writes. It currently
supports 2 simplifications, but more are planned.
It removes write accesses that write a loaded value back to the location
it was loaded from. It is a typical artifact from DeLICM. Removing it
will get rid of bogus dependencies later in dependency analysis.
It also removes statements without side-effects. ScopInfo already
removes these, but the removal of unnecessary writes can result in
more side-effect free statements.
Differential Revision: https://reviews.llvm.org/D30820
llvm-svn: 297473
In case LLVM pointers are annotated with !dereferencable attributes/metadata
or LLVM can look at the allocation from which a pointer is derived, we can know
that dereferencing pointers is safe and can be done unconditionally. We use this
information to proof certain pointers as save to hoist and then hoist them
unconditionally.
llvm-svn: 297375
One of the current limitations of DeLICM is that it only creates
PHI WRITEs that it knows are read by some PHI. Such writes may not span
all instances of a statement. Polly's code generator currently does not
support MemoryAccesses that are not executed in all instances
('partial accesses') and so has to give up on a possible mapping.
This workaround has once been suggested by Tobias Grosser: Try to
interpolate an arbitrary expansion to all instances. It will be checked
for possible conflicts with the existing Knowledge and can be applied if
the conflict checking result is that no semantics are changed.
Expansion is done by simplifying the mapping by coalescing with the hope
that coalescing will find a polyhedral 'rule' of the relevant map. It is
then 'gist'-ed using the domain of the relevant instances such that the
rule is expanded to the universe and finally intersected with the domain
of all statement instances.
The expansion makes conflicts become more likely, the found rule may
still not encompass all statement instances and the found rule exposes
internals of isl's implementation of coalesce and gist. The latter means
that the result depends on how much effort the implementation invests
into finding a rule which may change between versions of isl. Trivial
implementations of gist and coalesce just return the input arguments.
A patch that makes codegen support partial accesses is in preparation
as well.
Differential Revision: https://reviews.llvm.org/D30763
llvm-svn: 297373
Simplify ScopDetection::isInvariant(). Essentially deny everything that
is defined within the SCoP and is not load-hoisted.
The previous understanding of "invariant" has a few holes:
- Expressions without side-effects with only invariant arguments, but
are defined withing the SCoP's region with the exception of selects
and PHIs. These should be part of the index expression derived by
ScalarEvolution and not of the base pointer.
- Function calls with that are !mayHaveSideEffects() (typically
functions with "readnone nounwind" attributes). An example is given
below.
@C = external global i32
declare float* @getNextBasePtr(float*) readnone nounwind
...
%ptr = call float* @getNextBasePtr(float* %A, float %B)
The call might return:
* %A, so %ptr aliases with it in the SCoP
* %B, so %ptr aliases with it in the SCoP
* @C, so %ptr aliases with it in the SCoP
* a new pointer everytime it is called, such as malloc()
* a pointer into the allocated block of one of the aforementioned
* any of the above, at random at each call
Hence and contrast to a comment in the base_pointer.ll regression
test, %ptr is not necessarily the same all the time. It might also
alias with anything and no AliasAnalysis can tell otherwise if the
definition is external. It is hence not suitable in the role of a
base pointer.
The practical problem with base pointers defined in SCoP statements is
that it is not available globally in the SCoP. The statement instance
must be executed first before the base pointer can be used. This is no
problem if the base pointer is transferred as a scalar value between
statements. Uses of MemoryAccess::setNewAccessRelation may add a use of
the base pointer anywhere in the array. setNewAccessRelation is used by
JSONImporter, DeLICM and D28518. Indeed, BlockGenerator currently
assumes that base pointers are available globally and generates invalid
code for new access relation (referring to the base pointer of the
original code) if not, even if the base pointer would be available in
the statement.
This could be fixed with some added complexity and restrictions. The
ExprBuilder must lookup the local BBMap and code that call
setNewAccessRelation must check whether the base pointer is available
first.
The code would still be incorrect in the presence of aliasing. There
is the switch -polly-ignore-aliasing to explicitly allow this, but
it is hardly a justification for the additional complexity. It would
still be mostly useless because in most cases either getNextBasePtr()
has external linkage in which case the readnone nounwind attributes
cannot be derived in the translation unit itself, or is defined in the
same translation unit and gets inlined.
Reviewed By: grosser
Differential Revision: https://reviews.llvm.org/D30695
llvm-svn: 297281
Only when load-hoisted we can be sure the base pointer is invariant
during the SCoP's execution. Most of the time it would be added to
the required hoists for the alias checks anyway, except with
-polly-ignore-aliasing, -polly-use-runtime-alias-checks=0 or if
AliasAnalysis is already sure it doesn't alias with anything
(for instance if there is no other pointer to alias with).
Two more parts in Polly assume that this load-hoisting took place:
- setNewAccessRelation() which contains an assert which tests this.
- BlockGenerator which would use to the base ptr from the original
code if not load-hoisted (if the access expression is regenerated)
Differential Revision: https://reviews.llvm.org/D30694
llvm-svn: 297195
There is no point in optimizing unreachable code, hence our test cases should
always return.
This commit is part of a series that makes Polly more robust on the presence of
unreachables.
llvm-svn: 297158
These test cases should work in combination with
https://reviews.llvm.org/D12676, but became outdated over time. Update them
in preparation of discussions with Daniel Berlin on how to represent unreachable
in the post-dominator tree.
llvm-svn: 297157
Scops that exit with an unreachable are today still permitted, but make little
sense to optimize. We therefore can already skip them during scop detection.
This speeds up scop detection in certain cases and also ensures that bugpoint
does not introduce unreachables when reducing test cases.
In practice this change should have little impact, as the performance of
unreachable code is unlikely to matter.
This commit is part of a series that makes Polly more robust in the presence
of unreachables.
llvm-svn: 297151
There is no point in optimizing unreachable code, hence our test cases should
always return.
This commit is part of a series that makes Polly more robust on the presence of
unreachables.
llvm-svn: 297150
There is no point in optimizing unreachable code, hence our test cases should
always return.
This commit is part of a series that makes Polly more robust on the presence of
unreachables.
llvm-svn: 297147
r296992 made ScalarEvolution's CompareValueComplexity less aggressive,
and that broke the polly test being fixed in this change. This change
explicitly bumps CompareValueComplexity in said test case to make it
pass.
Can someone from the polly team please can give me an idea on if this
case is important enough to have
scalar-evolution-max-value-compare-depth be 3 by default?
llvm-svn: 296994
Some Polly ACC test cases fail without a working NVPTX backend. We explicitly
specify this dependence in REQUIRES. Alternatively, we could have only marked
polly-acc as supported in case the NVPTX backend is available, but as we might
use other backends in the future, this does not seem to be the best choice.
For this to work, we also need to make the 'targets_to_build' information
available.
Suggested-by: Michael Kruse <llvm@meinersbur.de>
llvm-svn: 296853
These loads cannot be savely hoisted as the condition guarding the
non-affine region cannot be duplicated to also protect the hoisted load
later on. Today they are dropped in ScopInfo. By checking for this early, we
do not even try to model them and possibly can still optimize smaller regions
not containing this specific required-invariant load.
llvm-svn: 296744
Multi-disjunct access maps can easily result in inbound assumptions which
explode in case of many memory accesses and many parameters. This change reduces
compilation time of some larger kernel from over 15 minutes to less than 16
seconds.
Interesting is the test case test/ScopInfo/multidim_param_in_subscript.ll
which has a memory access
[n] -> { Stmt_for_body3[i0, i1] -> MemRef_A[i0, -1 + n - i1] }
which requires folding, but where only a single disjunct remains. We can still
model this test case even when only using limited memory folding.
For people only reading commit messages, here the comment that explains what
memory folding is:
To recover memory accesses with array size parameters in the subscript
expression we post-process the delinearization results.
We would normally recover from an access A[exp0(i) * N + exp1(i)] into an
array A[][N] the 2D access A[exp0(i)][exp1(i)]. However, another valid
delinearization is A[exp0(i) - 1][exp1(i) + N] which - depending on the
range of exp1(i) - may be preferrable. Specifically, for cases where we
know exp1(i) is negative, we want to choose the latter expression.
As we commonly do not have any information about the range of exp1(i),
we do not choose one of the two options, but instead create a piecewise
access function that adds the (-1, N) offsets as soon as exp1(i) becomes
negative. For a 2D array such an access function is created by applying
the piecewise map:
[i,j] -> [i, j] : j >= 0
[i,j] -> [i-1, j+N] : j < 0
After this patch we generate only the first case, except for situations where
we can proove the first case to be invalid and can consequently select the
second without introducing disjuncts.
llvm-svn: 296679
Without this simplification for a loop nest:
void foo(long n1_a, long n1_b, long n1_c, long n1_d,
long p1_b, long p1_c, long p1_d,
float A_1[][p1_b][p1_c][p1_d]) {
for (long i = 0; i < n1_a; i++)
for (long j = 0; j < n1_b; j++)
for (long k = 0; k < n1_c; k++)
for (long l = 0; l < n1_d; l++)
A_1[i][j][k][l] += i + j + k + l;
}
the assumption:
n1_a <= 0 or (n1_a > 0 and n1_b <= 0) or
(n1_a > 0 and n1_b > 0 and n1_c <= 0) or
(n1_a > 0 and n1_b > 0 and n1_c > 0 and n1_d <= 0) or
(n1_a > 0 and n1_b > 0 and n1_c > 0 and n1_d > 0 and
p1_b >= n1_b and p1_c >= n1_c and p1_d >= n1_d)
is taken rather than the simpler assumption:
p9_b >= n9_b and p9_c >= n9_c and p9_d >= n9_d.
The former is less strict, as it allows arbitrary values of p1_* in case, the
loop is not executed at all. However, in practice these precise constraints
explode when combined across different accesses and loops. For now it seems
to make more sense to take less precise, but more scalable constraints by
default. In case we find a practical example where more precise constraints
are needed, we can think about allowing such precise constraints in specific
situations where they help.
This change speeds up the new test case from taking very long (waited at least
a minute, but it probably takes a lot more) to below a second.
llvm-svn: 296456
This patch adds an option to build against a version of libisl already
installed on the system. The installation is autodetected using the
pkg-config file shipped with isl.
The detection of the library is in the FindISL.cmake module that creates
an imported target.
Contributed-by: Philip Pfaffe <philip.pfaffe@gmail.com>
Differential Revision: https://reviews.llvm.org/D30043
llvm-svn: 296361
These verify that some scalars are not mapped because it would be
incorrect to do so.
For these check we verify that no transformation has been executed from
output of the pass's '-analyze'. Adding optimization remarks is not useful
as it would result in too many messages, even repeated ones. I avoided
checking the '-debug-only=polly-delicm' output which is an antipattern.
llvm-svn: 296348
Currently, pattern based optimizations of Polly can identify matrix
multiplication and optimize it according to BLIS matmul optimization pattern
(see ScheduleTreeOptimizer for details). This patch makes optimizations
based on pattern matching be enabled by default.
Reviewed-by: Tobias Grosser <tobias@grosser.es>
Differential Revision: https://reviews.llvm.org/D30293
llvm-svn: 295958
These tests were not included in the main DeLICM commit. These check the
cases where zone analysis cannot be successful because of assumption
violations.
We use the LLVM optimization remark infrastructure as it seems to be the
best fit for this kind of messages. I tried to make use if the
OptimizationRemarkEmitter. However, it would insert additional function
passes into the pass manager to get the hotness information. The pass
manager would insert them between the flatten pass and delicm, causing
the ScopInfo with the flattened schedule being thrown away.
Differential Revision: https://reviews.llvm.org/D30253
llvm-svn: 295846
Implement the -polly-delicm pass. The pass intends to undo the
effects of LoopInvariantCodeMotion (LICM) which adds additional scalar
dependencies into SCoPs. DeLICM will try to map those scalars back to
the array elements they were promoted from, as long as the array
element is unused.
The is the main patch from the DeLICM/DePRE patch series. It does not
yet undo GVN PRE for which additional information about known values
is needed and does not handle PHI write accesses that have have no
target. As such its usefulness is limited. Patches for these issues
including regression tests for error situatons will follow.
Reviewers: grosser
Differential Revision: https://reviews.llvm.org/D24716
llvm-svn: 295713
Instead of counting the number of read-only accesses, we now count the number of
distinct read-only array references when checking if a run-time alias check
may be too complex. The run-time alias check is quadratic in the number of
base pointers, not the number of accesses.
Before this change we accidentally skipped SPEC's lbm test case.
llvm-svn: 295567
This test case is a mini performance test case that shows the time needed for a
couple of simple reductions. It takes today about 325ms on my machine to run
this test case through 'opt' with scop construction and reduction detection. It
can be used as mini-proxy for further tuning of the reduction code.
Generally we do not commit performance test cases, but as this is very
small and also very fast it seems OK to keep it in the lit test suite.
This test case will also help to verify that future changes to the reduction
code will not affect the ordering of the reduction sets and will consequently
not cause spurious performance changes that only result from reordering of
dependences in the reduction set.
llvm-svn: 295549
Trying to fold such kind of dimensions will result in a division by zero,
which crashes the compiler. As such arrays are likely to invalidate the
scop anyhow (but are not illegal in LLVM-IR), there is no point in trying
to optimize the array layout. Hence, we just avoid the folding of
constant dimensions of size zero.
llvm-svn: 295415
Before this change wrapping range metadata resulted in exponential growth of
the context, which made context construction of large scops very slow. Instead,
we now just do not model the range information precisely, in case the number
of disjuncts in the context has already reached a certain limit.
llvm-svn: 295360
Commit r230230 introduced the use of range metadata to derive bounds for
parameters, instead of just looking at the type of the parameter. As part of
this commit support for wrapping ranges was added, where the lower bound of a
parameter is larger than the upper bound:
{ 255 < p || p < 0 }
However, at the same time, for wrapping ranges support for adding bounds given
by the size of the containing type has acidentally been dropped. As a result,
the range of the parameters was not guaranteed to be bounded any more. This
change makes sure we always add the bounds given by the size of the type and
then additionally add bounds based on signed wrapping, if available. For a
parameter p with a type size of 32 bit, the valid range is then:
{ -2147483648 <= p <= 2147483647 and (255 < p or p < 0) }
llvm-svn: 295349
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
Update test after commit r284501:
[SCEV] Make CompareValueComplexity a little bit smarter
Contributed-by: Sanjoy Das <sanjoy@playingwithpointers.com>
llvm-svn: 284543
lit recursively iterates through the test subdirectories and finds the ISL
unittest. For this test to work, the polly-isl-test executable needs to be
compiled.
Add the polly-isl-test dependency to POLLY_TEST_DEPS. This makes check-polly and
check-polly-tests work from a fresh build directory.
llvm-svn: 284339
The test non_affine_loop_used_later.ll also tests the profability heuristic. Add
the option -polly-unprofitable-scalar-accs explicitely to ensure that the test
succeeds if the default value is changed.
llvm-svn: 284338
Under some conditions MK_Value read accessed where converted to MK_ExitPHI read
accessed. This is unexpected because MK_ExitPHI read accesses are implicit after
the scop execution. This behaviour was introduced in r265261, which fixed a
failed assertion/crash in CodeGen.
Instead, we fix this failure in CodeGen itself. createExitPHINodeMerges(),
despite its name, also handles accesses of kind MK_Value, only to skip them
because they access values that are usually not PHI nodes in the SCoP region's
exit block. Except in the situation observed in r265261.
Do not convert value accessed to ExitPHI accesses and do not handle
value accesses like ExitPHI accessed in CodeGen anymore.
llvm-svn: 284023
Folders in Visual Studio solutions help organize the build artifacts from all
LLVM projects. There is a folder to keep Polly-built files in.
llvm-svn: 283546
Running isl tests is important to gain confidence that the isl build we created
works as expected. Besides the actual isl tests, there are also isl AST
generation tests shipped with isl. This change only adds support for the isl
unit tests. AST generation test support is left for a later commit.
There is a choice to run tests directly through the build system or in the
context of lit. We choose to run tests as part of lit to as this allows us to
easily set environment variables, print output only on error and generally run
the tests directly from the lit command.
Reviewers: brad.king, Meinersbur
Subscribers: modocache, brad.king, pollydev, beanz, llvm-commits, mgorny
Differential Revision: https://reviews.llvm.org/D25155
llvm-svn: 283245
With this option one can disable the heuristic that assumes that statements with
a scalar write access cannot be profitably optimized. Such a statement instances
necessarily have WAW-dependences to itself. With DeLICM scalar accesses can be
changed to array accesses, which can avoid these WAW-dependence.
llvm-svn: 283233
ScopArrayInfo used to determine base pointer origins by looking up whether the
base pointer is a load. The "base pointer" for scalar accesses is the
llvm::Value being accessed. This is only a symbolic base pointer, it
represents the alloca variable (.s2a or .phiops) generated for it at code
generation.
This patch disables determining base pointer origin for scalars.
A test case where this caused a crash will be added in the next commit. In that
test SAI tried to get the origin base pointer that was only declared later,
therefore not existing. This is probably only possible for scalars used in
PHINode incoming blocks.
llvm-svn: 283232
Summary:
Both `canUseISLTripCount()` and `addOverApproximatedRegion()` contained checks
to reject endless loops which are now removed and replaced by a single check
in `isValidLoop()`.
For reporting such loops the `ReportLoopOverlapWithNonAffineSubRegion` is
renamed to `ReportLoopHasNoExit`. The test case
`ReportLoopOverlapWithNonAffineSubRegion.ll` is adapted and renamed as well.
The schedule generation in `buildSchedule()` is based on the following
assumption:
Given some block B that is contained in a loop L and a SESE region R,
we assume that L is contained in R or the other way around.
However, this assumption is broken in the presence of endless loops that are
nested inside other loops. Therefore, in order to prevent erroneous behavior
in `buildSchedule()`, r265280 introduced a corresponding check in
`canUseISLTripCount()` to reject endless loops. Unfortunately, it was possible
to bypass this check with -polly-allow-nonaffine-loops which was fixed by adding
another check to reject endless loops in `allowOverApproximatedRegion()` in
r273905. Hence there existed two separate locations that handled this case.
Thank you Johannes Doerfert for helping to provide the above background
information.
Reviewers: Meinersbur, grosser
Subscribers: _jdoerfert, pollydev
Differential Revision: https://reviews.llvm.org/D24560
Contributed-by: Matthias Reisinger <d412vv1n@gmail.com>
llvm-svn: 281987
In case sequential kernels are found deeper in the loop tree than any parallel
kernel, the overall scop is probably mostly sequential. Hence, run it on the
CPU.
llvm-svn: 281849
Offloading to a GPU is only beneficial if there is a sufficient amount of
compute that can be accelerated. Many kernels just have a very small number
of dynamic compute, which means GPU acceleration is not beneficial. We
compute at run-time an approximation of how many dynamic instructions will be
executed and fall back to CPU code in case this number is not sufficiently
large. To keep the run-time checking code simple, we over-approximate the
number of instructions executed in each statement by computing the volume of
the rectangular hull of its iteration space.
llvm-svn: 281848
We may generate GPU kernels that store into scalars in case we run some
sequential code on the GPU because the remaining data is expected to already be
on the GPU. For these kernels it is important to not keep the scalar values
in thread-local registers, but to store them back to the corresponding device
memory objects that backs them up.
We currently only store scalars back at the end of a kernel. This is only
correct if precisely one thread is executed. In case more than one thread may
be run, we currently invalidate the scop. To support such cases correctly,
we would need to always load and store back from a corresponding global
memory slot instead of a thread-local alloca slot.
llvm-svn: 281838
Our alias checks precisely check that the minimal and maximal accessed elements
do not overlap in a kernel. Hence, we must ensure that our host <-> device
transfers do not touch additional memory locations that are not covered in
the alias check. To ensure this, we make sure that the data we copy for a
given array is only the data from the smallest element accessed to the largest
element accessed.
We also adjust the size of the array according to the offset at which the array
is actually accessed.
An interesting result of this is: In case array are accessed with negative
subscripts ,e.g., A[-100], we automatically allocate and transfer _more_ data to
cover the full array. This is important as such code indeed exists in the wild.
llvm-svn: 281611
This is the fourth patch to apply the BLIS matmul optimization pattern on matmul
kernels (http://www.cs.utexas.edu/users/flame/pubs/TOMS-BLIS-Analytical.pdf).
BLIS implements gemm as three nested loops around a macro-kernel, plus two
packing routines. The macro-kernel is implemented in terms of two additional
loops around a micro-kernel. The micro-kernel is a loop around a rank-1
(i.e., outer product) update. In this change we perform copying to created
arrays, which is the last step to implement the packing transformation.
Reviewed-by: Tobias Grosser <tobias@grosser.es>
Differential Revision: https://reviews.llvm.org/D23260
llvm-svn: 281441
We do not need the size of the outermost dimension in most cases, but if we
allocate memory for newly created arrays, that size is needed.
Reviewed-by: Michael Kruse <llvm@meinersbur.de>
Differential Revision: https://reviews.llvm.org/D23991
llvm-svn: 281234
Instead of aborting, we now bail out gracefully in case the kernel IR we
generate is invalid. This can currently happen in case the SCoP stores
pointer values, which we model as arrays, as data values into other arrays. In
this case, the original pointer value is not available on the device and can
consequently not be stored. As detecting this ahead of time is not so easy, we
detect these situations after the invalid IR has been generated and bail out.
llvm-svn: 281193
If these arrays have never been accessed we failed to derive an upper bound
of the accesses and consequently a size for the outermost dimension. We
now explicitly check for empty access sets and then just use zero as size
for the outermost dimension.
llvm-svn: 281165
The -polly-flatten-schedule pass reduces the number of scattering
dimensions in its isl_union_map form to make them easier to understand.
It is not meant to be used in production, only for debugging and
regression tests.
To illustrate, how it can make sets simpler, here is a lifetime set
used computed by the porposed DeLICM pass without flattening:
{ Stmt_reduction_for[0, 4] -> [0, 2, o2, o3] : o2 < 0;
Stmt_reduction_for[0, 4] -> [0, 1, o2, o3] : o2 >= 5;
Stmt_reduction_for[0, 4] -> [0, 1, 4, o3] : o3 > 0;
Stmt_reduction_for[0, i1] -> [0, 1, i1, 1] : 0 <= i1 <= 3;
Stmt_reduction_for[0, 4] -> [0, 2, 0, o3] : o3 <= 0 }
And here the same lifetime for a semantically identical one-dimensional
schedule:
{ Stmt_reduction_for[0, i1] -> [2 + 3i1] : 0 <= i1 <= 4 }
Differential Revision: https://reviews.llvm.org/D24310
llvm-svn: 280948
The check-polly-tests target runs regression/unit tests but without checking
formatting. This is useful to not having to reload a file in an open editor
(which eg. clears the undo buffer, moves cursor/window position) when running
polly-update-format.
After this change, the following test targets exist:
- check-polly-unittests to run unittests only
- check-polly-tests to run unit and regression tests
- polly-check-format to check formatting using clang-format
- check-polly to run them all
As a side-effect, when running check-polly, polly-check-format and run in
parallel (instead of polly-check-format first).
Differential Revision: https://reviews.llvm.org/D24191
llvm-svn: 280654
Change the code around setNewAccessRelation to allow to use a an existing array
element for memory instead of an ad-hoc alloca. This facility will be used for
DeLICM/DeGVN to convert scalar dependencies into regular ones.
The changes necessary include:
- Make the code generator use the implicit locations instead of the alloca ones.
- A test case
- Make the JScop importer accept changes of scalar accesses for that test case.
- Adapt the MemoryAccess interface to the fact that the MemoryKind can change.
They are named (get|is)OriginalXXX() to get the status of the memory access
before any change by setNewAccessRelation() (some properties such as
getIncoming() do not change even if the kind is changed and are still
required). To get the modified properties, there is (get|is)LatestXXX(). The
old accessors without Original|Latest become synonyms of the
(get|is)OriginalXXX() to not make functional changes in unrelated code.
Differential Revision: https://reviews.llvm.org/D23962
llvm-svn: 280408
Add the infrastructure for unittests to Polly and two simple tests for
conversion between isl_val and APInt. In addition, a build target
check-polly-unittests is added to run only the unittests but not the regression
tests.
Clang's unittest mechanism served as as a blueprint which then was adapted to
Polly.
Differential Revision: https://reviews.llvm.org/D23833
llvm-svn: 279734
configure_lit_site_cfg defines some more parameters that are used in
lit.site.cfg.in. configure_file would leave those empty. These additional
definitions seem to be unimportant for regression tests, but unittests do not
work without them.
In case of out-of-tree builds, define the additional parameters with default
values. These may not take all configuration parameters into account, as
configure_lit_site_cfg would.
llvm-svn: 279733
Dump polyhedral descriptions of Scops optimized with the isl scheduling
optimizer and the set of post-scheduling transformations applied
on the schedule tree to be able to check the work of the IslScheduleOptimizer
pass at the polyhedral level.
Reviewed-by: Tobias Grosser <tobias@grosser.es>
Differential Revision: https://reviews.llvm.org/D23740
llvm-svn: 279395
The existing code would add the operands in the wrong order, and eventually
crash because the SCEV expression doesn't exactly match the parameter SCEV
expression in SCEVAffinator::visit. (SCEV doesn't sort the operands to
getMulExpr in general.)
Differential Revision: https://reviews.llvm.org/D23592
llvm-svn: 279087
We already invalidated a couple of critical values earlier on, but we now
invalidate all instructions contained in a scop after the scop has been code
generated. This is necessary as later scops may otherwise obtain SCEV
expressions that reference values in the earlier scop that before dominated
the later scop, but which had been moved into the conditional branch and
consequently do not dominate the later scop any more. If these very values are
then used during code generation of the later scop, we generate used that are
dominated by the values they use.
This fixes: http://llvm.org/PR28984
llvm-svn: 279047
Normally this is ensured when adding PHI nodes, but as PHI node dependences
do not need to be added in case all incoming blocks are within the same
non-affine region, this was missed.
This corrects an issue visible in LNT's sqlite3, in case invariant load hoisting
was disabled.
llvm-svn: 278792
This will make it easier to switch the default of Polly's invariant load
hoisting strategy and also makes it very clear that these test cases
indeed require invariant code hoisting to work.
llvm-svn: 278667
This is the third patch to apply the BLIS matmul optimization pattern on matmul
kernels (http://www.cs.utexas.edu/users/flame/pubs/TOMS-BLIS-Analytical.pdf).
BLIS implements gemm as three nested loops around a macro-kernel, plus two
packing routines. The macro-kernel is implemented in terms of two additional
loops around a micro-kernel. The micro-kernel is a loop around a rank-1
(i.e., outer product) update. In this change we perform replacement of
the access relations and create empty arrays, which are steps to implement
the packing transformation. In subsequent changes we will implement copying
to created arrays.
Reviewed-by: Tobias Grosser <tobias@grosser.es>
Differential Revision: http://reviews.llvm.org/D22187
llvm-svn: 278666
To do so we change the way array exents are computed. Instead of the precise
set of memory locations accessed, we now compute the extent as the range between
minimal and maximal address in the first dimension and the full extent defined
by the sizes of the inner array dimensions.
We also move the computation of the may_persist region after the construction
of the arrays, as it relies on array information. Without arrays being
constructed no useful information is computed at all.
llvm-svn: 278212
Ensure the right scalar allocations are used as the host location of data
transfers. For the device code, we clear the allocation cache before device
code generation to be able to generate new device-specific allocation and
we need to make sure to add back the old host allocations as soon as the
device code generation is finished.
llvm-svn: 278126
This increases the readability of the IR and also clarifies that the GPU
inititialization is executed _after_ the scalar initialization which needs
to before the code of the transformed scop is executed.
Besides increased readability, the IR should not change. Specifically, I
do not expect any changes in program semantics due to this patch.
llvm-svn: 278125
In case some code -- not guarded by control flow -- would be emitted directly in
the start block, it may happen that this code would use uninitalized scalar
values if the scalar initialization is only emitted at the end of the start
block. This is not a problem today in normal Polly, as all statements are
emitted in their own basic blocks, but Polly-ACC emits host-to-device copy
statements into the start block.
Additional Polly-ACC test coverage will be added in subsequent changes that
improve the handling of PHI nodes in Polly-ACC.
llvm-svn: 278124
After having generated the code for a ScopStmt, we run a simple dead-code
elimination that drops all instructions that are known to be and remain unused.
Until this change, we only considered instructions for dead-code elimination, if
they have a corresponding instruction in the original BB that belongs to
ScopStmt. However, when generating code we do not only copy code from the BB
belonging to a ScopStmt, but also generate code for operands referenced from BB.
After this change, we now also considers code for dead code elimination, which
does not have a corresponding instruction in BB.
This fixes a bug in Polly-ACC where such dead-code referenced CPU code from
within a GPU kernel, which is possible as we do not guarantee that all variables
that are used in known-dead-code are moved to the GPU.
llvm-svn: 278103
When adding code that avoids to pass values used in isl expressions and
LLVM instructions twice, we forgot to make single variable passed to the
kernel available in the ValueMap that makes it usable for instructions that
are not replaced with isl ast expressions. This change adds the variable
that is passed to the kernel to the ValueMap to ensure it is available
for such use cases as well.
llvm-svn: 278039
Before this commit we generated the array type in reverse order and we also
added the outermost dimension size to the new array declaration, which is
incorrect as Polly additionally assumed an additional unsized outermost
dimension, such that we had an off-by-one error in the linearization of access
expressions.
llvm-svn: 277802
These annotations ensure that the NVIDIA PTX assembler limits the number of
registers used such that we can be certain the resulting kernel can be executed
for the number of threads in a thread block that we are planning to use.
llvm-svn: 277799
Pass the content of scalar array references to the alloca on the kernel side
and do not pass them additional as normal LLVM scalar value.
llvm-svn: 277699
Otherwise, we would try to re-optimize them with Polly-ACC and possibly even
generate kernels that try to offload themselves, which does not work as the
GPURuntime is not available on the accelerator and also does not make any
sense.
llvm-svn: 277589
Extend the jscop interface to allow the user to export arrays. It is required
that already existing arrays of the list of arrays correspond to arrays
of the SCoP. Each array that is appended to the list will be newly created.
Furthermore, we allow the user to modify access expressions to reference
any array in case it has the same element type.
Reviewed-by: Tobias Grosser <tobias@grosser.es>
Differential Revision: https://reviews.llvm.org/D22828
llvm-svn: 277263
Before this change we used the array index, which would result in us accessing
the parameter array out-of-bounds. This bug was visible for test cases where not
all arrays in a scop are passed to a given kernel.
llvm-svn: 276961
Adding a new pass PolyhedralInfo. This pass will be the interface to Polly.
Initially, we will provide the following interface:
- #IsParallel(Loop *L) - return a bool depending on whether the loop is
parallel or not for the given program order.
Patch by Utpal Bora <cs14mtech11017@iith.ac.in>
Differential Revision: https://reviews.llvm.org/D21486
llvm-svn: 276637
Also factor out getArraySize() to avoid code dupliciation and reorder some
function arguments to indicate the direction into which data is transferred.
llvm-svn: 276636
At the beginning of each SCoP, we allocate device arrays for all arrays
used on the GPU and we free such arrays after the SCoP has been executed.
llvm-svn: 276635
Do not process SCoPs with infeasible runtime context in the new
ScopInfoWrapperPass. Do not compute dependences for such SCoPs in the new
DependenceInfoWrapperPass.
Patch by Utpal Bora <cs14mtech11017@iith.ac.in>
Differential Revision: https://reviews.llvm.org/D22402
llvm-svn: 276631
This is the second patch to apply the BLIS matmul optimization pattern
on matmul kernels
(http://www.cs.utexas.edu/users/flame/pubs/TOMS-BLIS-Analytical.pdf).
BLIS implements gemm as three nested loops around a macro-kernel, plus
two packing routines. The macro-kernel is implemented in terms
of two additional loops around a micro-kernel. The micro-kernel
is a loop around a rank-1 (i.e., outer product) update. In this change
we create the BLIS macro-kernel by applying a combination of tiling
and interchanging. In subsequent changes we will implement the packing
transformation.
Reviewed-by: Tobias Grosser <tobias@grosser.es>
Differential Revision: http://reviews.llvm.org/D21491
llvm-svn: 276627
There is no need to expose the selected device at the moment. We also pass back
pointers as return values, as this simplifies the interface.
llvm-svn: 276623
Run the NVPTX backend over the GPUModule IR and write the resulting assembly
code in a string.
To work correctly, it is important to invalidate analysis results that still
reference the IR in the kernel module. Hence, this change clears all references
to dominators, loop info, and scalar evolution.
Finally, the NVPTX backend has troubles to generate code for various special
floating point types (not surprising), but also for uncommon integer types. This
commit does not resolve these issues, but pulls out problematic test cases into
separate files to XFAIL them individually and resolve them in future (not
immediate) changes one by one.
llvm-svn: 276396
This change introduces the actual compute code in the GPU kernels. To ensure
all values referenced from the statements in the GPU kernel are indeed available
we scan all ScopStmts in the GPU kernel for references to llvm::Values that
are not yet covered by already modeled outer loop iterators, parameters, or
array base pointers and also pass these additional llvm::Values to the
GPU kernel.
For arrays used in the GPU kernel we introduce a new ScopArrayInfo object, which
is referenced by the newly generated access functions within the GPU kernel and
which is used to help with code generation.
llvm-svn: 276270
This ensures that no trivially dead code is generated. This is not only cleaner,
but also avoids troubles in case code is generated in a separate function and
some of this dead code contains references to values that are not available.
This issue may happen, in case the memory access functions have been updated
and old getelementptr instructions remain in the code. With normal Polly,
a test case is difficult to draft, but the upcoming GPU code generation can
possibly trigger such problems. We will later extend this dead-code elimination
to region and vector statements.
llvm-svn: 276263
It seems the order in which we generated memory accesses changed such that
the import of these updated memory accesses failed for the 'loop3' statement
in this test case. Unfortunately, the existing CHECK lines were not strict
enough to catch this. Hence, besides fixing the order of the memory access
lines we also ensure that the memory access changes are both clearly visibly
and well checked.
llvm-svn: 276247
This is currently not supported and will only be added later. Also update the
test cases to ensure no invariant code hoisting is applied.
llvm-svn: 275987
We use this opportunity to further classify the different user statements that
can arise and add TODOs for the ones not yet implemented.
llvm-svn: 275957
Create for each kernel a separate LLVM-IR module containing a single function
marked as kernel function and taking one pointer for each array referenced
by this kernel. Add debugging output to verify the kernels are generated
correctly.
llvm-svn: 275952
Initialize the list of references to a GPU array to ensure that the arrays that
need to be passed to kernel calls are computed correctly. Furthermore, the very
same information is also necessary to compute synchronization correctly. As the
functionality to compute these references is already available, what is left for
us to do is only to connect the necessary functionality to compute array
reference information.
llvm-svn: 275798
Create LLVM-IR for all host-side control flow of a given GPU AST. We implement
this by introducing a new GPUNodeBuilder class derived from IslNodeBuilder. The
IslNodeBuilder will take care of generating all general-purpose ast nodes, but
we provide our own createUser implementation to handle the different GPU
specific user statements. For now, we just skip any user statement and only
generate a host-code sceleton, but in subsequent commits we will add handling of
normal ScopStmt's performing computations, kernel calls, as well as host-device
data transfers. We will also introduce run-time check generation and LICM in
subsequent commits.
llvm-svn: 275783
Otherwise ppcg would try to call into pet functionality that this not available,
which obviously will cause trouble. As we can easily print these statements
ourselves, we just do so.
llvm-svn: 275579
This option increases the scalability of the scheduler and allows us to remove
the 'gisting' workaround we introduced in r275565 to handle a more complicated
test case. Another benefit of using this option is also that the generated
code looks a lot more streamlined.
Thanks to Sven Verdoolaege for reminding me of this option.
llvm-svn: 275573
This works around a shortcoming of the isl scheduler, which even for some
smaller test cases does not terminate in case domain constraints are part
of the flow dependences.
llvm-svn: 275565
Arrays with integer base type are similar to arrays with floating point types,
with the exception that LLVM's integer types can take some odd values. We
add a selection of different values to make sure we correctly round these
types when necessary.
References to scalar integer types are special, as we currently do not model
these types as array accesses as they are considered 'synthesizable' by Polly.
As a result, we do not generate explicit data-transfers for them, but instead
will need to keep track of all references to 'synthesizable' values separately.
At the current stage, this is only visible by missing host-to-device
data-transfer calls. In the future, we will also require special code generation
strategies.
llvm-svn: 275551
We currently only test that the code structure we generate for these scalar
parameters is correct and we add these types to make sure later code generation
additions have sufficient test coverage.
In case some of these types cannot be mapped due to missing hardware support
on the GPU some of these test cases may need to be updated later on.
llvm-svn: 275548
A sequence of CHECK lines allows additional statements to appear in the
output of the tested program without any test failures appearing. As we do
not want this to happen, switch this test case to use CHECK-NEXT.
llvm-svn: 275534
For this we need to provide an explicit list of statements as they occur in
the polly::Scop to ppcg.
We also setup basic AST printing facilities to facilitate debugging. To allow
code reuse some (minor) changes in ppcg are have been necessary.
llvm-svn: 275436
The tile size was previously uninitialized. As a result, it was often zero (aka.
no tiling), which is not what we want in general. More importantly, there was
the risk for arbitrary tile sizes to be choosen, which we did not observe, but
which still is highly problematic.
llvm-svn: 275418
This change now applies ppcg's GPU mapping on our initial schedule. For this
to work, we need to also initialize the set of all names (isl_ids) used in
the scop as well as the program context.
llvm-svn: 275396
To do so we copy the necessary information to compute an initial schedule from
polly::Scop to ppcg's scop. Most of the necessary information is directly
available and only needs to be passed on to ppcg, with the exception of 'tagged'
access relations, access relations that additionally carry information about
which memory access an access relation originates from.
We could possibly perform the construction of tagged accesses as part of
ScopInfo, but as this format is currently specific to ppcg we do not do this
yet, but keep this functionality local to our GPU code generation.
After the scop has been initialized, we compute data dependences and ask ppcg to
compute an initial schedule. Some of this functionality is already available in
polly::DependenceInfo and polly::ScheduleOptimizer, but to keep differences
to ppcg small we use ppcg's functionality here. We may later investiage if
a closer integration of these tools makes sense.
llvm-svn: 275390
At this stage, we do not yet modify the IR but just generate a default
initialized ppcg_scop and gpu_prog and free both immediately. Both will later be
filled with data from the polly::Scop and are needed to use PPCG for GPU
schedule generation. This commit does not yet perform any GPU code generation,
but ensures that the basic infrastructure has been put in place.
We also add a simple test case to ensure the new code is run and use this
opportunity to verify that GPU_CODEGEN tests are only run if GPU code generation
has been enabled in cmake.
llvm-svn: 275389
Check not only that the compiler is not crashing, but also whether the
probablematic part (The sequence of instructions simplified to '4') is reflected
in the output.
Thanks to Tobias for the hint.
llvm-svn: 275189
An assertion in visitSDivInstruction() checked whether the divisor is constant
by checking whether the argument is a ConstantInt. However, SCEVValidator allows
the divisor to be simplified to a constant by ScalarEvolution.
We synchronize the implementation of SCEVValidator and SCEVAffinator to both
accept simplified SCEV expressions.
llvm-svn: 275174
For llvm the memory accesses from nonaffine loops should be visible,
however for polly those nonaffine loops should be invisible/boxed.
This fixes llvm.org/PR28245
Cointributed-by: Huihui Zhang <huihuiz@codeaurora.org>
Differential Revision: http://reviews.llvm.org/D21591
llvm-svn: 274842
Reject and report regions that contains loops overlapping nonaffine region.
This situation typically happens in the presence of inifinite loops.
This addresses bug llvm.org/PR28071.
Differential Revision: http://reviews.llvm.org/D21312
Contributed-by: Huihui Zhang <huihuiz@codeaurora.org>
llvm-svn: 273905
This patch addresses:
- A new function pass to compute polyhedral dependences. This is
required to avoid the region pass manager.
- Stores a map of Scop to Dependence object for all the scops present
in a function. By default, access wise dependences are stored.
Patch by Utpal Bora <cs14mtech11017@iith.ac.in>
Differential Revision: http://reviews.llvm.org/D21105
llvm-svn: 273881
This patch adds a new function pass ScopInfoWrapperPass so that the
polyhedral description of a region, the SCoP, can be constructed and
used in a function pass.
Patch by Utpal Bora <cs14mtech11017@iith.ac.in>
Differential Revision: http://reviews.llvm.org/D20962
llvm-svn: 273856
This is the first patch to apply the BLIS matmul optimization pattern
on matmul kernels
(http://www.cs.utexas.edu/users/flame/pubs/TOMS-BLIS-Analytical.pdf).
BLIS implements gemm as three nested loops around a macro-kernel,
plus two packing routines. The macro-kernel is implemented in terms
of two additional loops around a micro-kernel. The micro-kernel
is a loop around a rank-1 (i.e., outer product) update.
In this change we create the BLIS micro-kernel by applying
a combination of tiling and unrolling. In subsequent changes
we will add the extraction of the BLIS macro-kernel
and implement the packing transformation.
Contributed-by: Roman Gareev <gareevroman@gmail.com>
Reviewed-by: Tobias Grosser <tobias@grosser.es>
Differential Revision: http://reviews.llvm.org/D21140
llvm-svn: 273397
With this update the isl AST generation extracts disjunctive constraints early
on. As a result, code that previously resulted in two branches with (close-to)
identical code within them:
if (P <= -1) {
for (int c0 = 0; c0 < N; c0 += 1)
Stmt_store(c0);
} else if (P >= 1)
for (int c0 = 0; c0 < N; c0 += 1)
Stmt_store(c0);
results now in only a single branch body:
if (P <= -1 || P >= 1)
for (int c0 = 0; c0 < N; c0 += 1)
Stmt_store(c0);
This resolves http://llvm.org/PR27559
Besides the above change, this isl update brings better simplification of
sets/maps containing existentially quantified dimensions and fixes a bug in
isl's coalescing.
llvm-svn: 272500
As these test cases will be changed in a subsequent commit, we expand and
tighten them to make the subsequent changes to them more obvious. As part of
this we add more context to some test cases and add CHECK-NEXT lines to ensure
no intermediate lines are missed by accident.
llvm-svn: 272499
IntToPtr and PtrToInt instructions are basically no-ops that we can handle as
such. In order to generate them properly as parameters we had to improve the
ScopExpander, though the change is the first in the direction of a more
aggressive scalar synthetization.
This patch was originally contributed by Johannes Doerfert in r271888, but was
in conflict with the revert in r272483. This is a recommit with some minor
adjustment to the test cases to take care of differing instruction names.
llvm-svn: 272485
The recent expression type changes still need more discussion, which will happen
on phabricator or on the mailing list. The precise list of commits reverted are:
- "Refactor division generation code"
- "[NFC] Generate runtime checks after the SCoP"
- "[FIX] Determine insertion point during SCEV expansion"
- "Look through IntToPtr & PtrToInt instructions"
- "Use minimal types for generated expressions"
- "Temporarily promote values to i64 again"
- "[NFC] Avoid unnecessary comparison for min/max expressions"
- "[Polly] Fix -Wunused-variable warnings (NFC)"
- "[NFC] Simplify min/max expression generation"
- "Simplify the type adjustment in the IslExprBuilder"
Some of them are just reverted as we would otherwise get conflicts. I will try
to re-commit them if possible.
llvm-svn: 272483
Siddharth Bhat