Add the options -polly-codegen-trace-stmts and
-polly-codegen-trace-scalars. When enabled, adds a call to the
beginning of every generated statement that prints the executed
statement instance. With -polly-codegen-trace-scalars, it also prints
the value of all scalars that are used in the statement, and PHIs
defined in the beginning of the statement.
Differential Revision: https://reviews.llvm.org/D45743
llvm-svn: 330864
The current statement domain derivation algorithm does not (always)
consider that different exit blocks of a loop can have different
conditions to be reached.
From the code
for (int i = n; ; i-=2) {
if (i <= 0) goto even;
if (i <= 1) goto odd;
A[i] = i;
}
even:
A[0] = 42;
return;
odd:
A[1] = 21;
return;
Polly currently derives the following domains:
Stmt_even_critedge
Domain :=
[n] -> { Stmt_even_critedge[] };
Stmt_odd
Domain :=
[n] -> { Stmt_odd[] : (1 + n) mod 2 = 0 and n > 0 };
while the domain for the odd case is correct, Stmt_even is assumed to be
executed unconditionally, which is obviously wrong. While projecting out
the loop dimension in `adjustDomainDimensions`, it does not consider
that there are other exit condition that have matched before.
I don't know a how to fix this without changing a lot of code. Therefore
This patch rejects loops with multiple exist blocks to fix the
miscompile of test-suite's uuencode.
The odd condition is transformed by LLVM to
%cmp1 = icmp eq i64 %indvars.iv, 1
such that the project_out in adjustDomainDimensions() indeed only
matches for odd n (using this condition only, we'd have an infinite loop
otherwise).
The even condition manifests as
%cmp = icmp slt i64 %indvars.iv, 3
Because buildDomainsWithBranchConstraints() does not consider other exit
conditions, it has to assume that the induction variable will eventually
be lower than 3 and taking this exit.
IMHO we need to reuse the algorithm that determines the number of
iterations (addLoopBoundsToHeaderDomain) to determine which exit
condition applies first. It has to happen in
buildDomainsWithBranchConstraints() because the result will need to
propagate to successor BBs. Currently addLoopBoundsToHeaderDomain() just
look for union of all backedge conditions (which means leaving not the
loop here). The patch in llvm.org/PR35465 changes it to look for exit
conditions instead. This is required because there might be other exit
conditions that do not alternatively go back to the loop header.
Differential Revision: https://reviews.llvm.org/D45649
llvm-svn: 330858
Before this patch, ISL_ASSERT only printed an error message to stderr.
This can be easily missed if the program continues or just fails later.
To fail-early and help error diagnostics (e.g. using bugpoint), call
abort() when an assertion does not hold.
I seem to just have forgotten to add this abort() when I originally
proposed the ISL_ASSERT macro.
Suggested-By: Eli Friedman <efriedma@codeaurora.org>
Differential Revision: https://reviews.llvm.org/D45171
llvm-svn: 330467
Add the switch -polly-debug-func to define the name of a debug
function. This function is ignored for any validity check.
Its purpose is to allow to observe a value after transformation by a
SCoP, and to follow which statements are executed in which order. For
instance, consider the following code:
static void dbg_printf(int sum, int i) {
fprintf(stderr, "The value of sum is %d, i=%d\n", sum, i);
fflush(stderr);
}
void func(int n) {
int sum = 0;
for (int i = 0; i < 16; i+=1) {
sum += i;
dbg_printf(sum, i);
}
}
Executing this after Polly's codegen with -polly-debug-func=dbg_printf
reveals the new execution order and the assumed values at that point of
execution.
Differential Revision: https://reviews.llvm.org/D45728
llvm-svn: 330466
In r330292 this assert was turned incorrectly into an unreachable, but
the correct behavior (thanks Michael) is to assert for anything that is
not 64 bit, but falltrough for 64 bit. I document this in the source
code.
llvm-svn: 330309
Originally the RuntimeDebugBuilder printed vectors with withspaces
between the elements. This historic use is meanwhile gone, but the
functionality is still available.
We now change the behavior to print elements just one after the other
without adding white spaces in between. This is useful for D45743, an
upcoming commmit, which also adds test coverage for this feature.
In general, printing elements of a vector directly is more generic as
it allows uses where no white-spaces are desired. Specifically, it
allows the user to build vectors of items to be printed where their
length is only known at run-time.
llvm-svn: 330292
Summary:
As of rL329273, LLVM has a mechanism to load new-pm plugins in opt. Use
this API in Polly.
Reviewers: grosser, Meinersbur, bollu
Reviewed By: grosser, Meinersbur
Subscribers: lksbhm, bollu, pollydev, llvm-commits
Differential Revision: https://reviews.llvm.org/D45484
llvm-svn: 330181
Piecewise affine expressions have directly corresponding mathematical
operators. Introduce these operators as overloads as this makes writing
code with isl::pw_aff expressions more directly readable.
We can now write:
A = B + C instead of A = B.add(C)
Reviewers: Meinersbur, bollu, sebpop
Reviewed By: Meinersbur
Subscribers: philip.pfaffe, pollydev, llvm-commits
Differential Revision: https://reviews.llvm.org/D45534
llvm-svn: 329880
A check in assert-builds was meant to verify that a load provides a
value in all statement instances (i.e. its domain). The domain is
commonly gist'ed within the parameter context to contain fewer
constraints. However, statement instances outside the context are
no valid executions, hence the value provided can be undefined.
Refine the check for valid loads to only needed to be defined within
the SCoP context.
In addition, the JSONImporter had to be changed to allow importing
access relations that are broader than the current access relation,
but still defined over all statement instances.
This should fix the compiler crash in test-suite's oggenc of the
-polly-process-unprofitable buildbot.
llvm-svn: 329655
Commit r329640 introduced the removal of all MemoryAccesses of a Scop.
It accidentally continued iterating over a vector whose iterators
have been invalidated by a MemoryAccess removal.
Make a copy of the MemoryAccesses to remove to iterate over while
removing them.
llvm-svn: 329653
Removing a statement left its MemoryAccesses in some lists and maps of
the SCoP. Which lists depends on at which phase of the SCoP
construction the statement is deleted. Follow-up passes could still see
the already deleted MemoryAccesses by iterating through these
lists/maps, resulting in an access violation.
When removing a ScopStmt, also remove all its MemoryAccesses by using
the same mechnism that removes a MemoryAccess.
llvm-svn: 329640
std::remove, despite its name, does not remove elements from a list, but
only moves them to the end of a list. Call erase() to shorten the
vector to the remaining elements.
Test case included in next commit.
llvm-svn: 329639
This patch removes the heuristic in
- Polly :: lib/Support/ScopHelper.cpp
The heuristic forces blocks that directly follow a loop header to not to be considered error blocks.
It was introduced in r249611 with the following commit message:
> This replaces the support for user defined error functions by a
> heuristic that tries to determine if a call to a non-pure function
> should be considered "an error". If so the block is assumed not to be
> executed at runtime. While treating all non-pure function calls as
> errors will allow a lot more regions to be analyzed, it will also
> cause us to dismiss a lot again due to an infeasible runtime context.
> This patch tries to limit that effect. A non-pure function call is
> considered an error if it is executed only in conditionally with
> regards to a cheap but simple heuristic.
In the code below `CCK_Abort2()` would be considered as an error block, but not `CCK_Abort1()` due to this heuristic.
```
for (int i = 0; i < n; i+=1) {
if (ErrorCondition1)
CCK_Abort1(); // No __attribute__((noreturn))
if (ErrorCondition2)
CCK_Abort2(); // No __attribute__((noreturn))
}
```
This does not seem useful. Checking error conditions in the beginning of some work is quite common. It causes a switch default-case to be not considered an error block in SPEC's cactuBSSN. The comment justifying the heuristic mentions a "load", which does not seem to be applicable here. It has been proposed to remove the heuristic.
In addition, the patch fixes the following test cases:
- Polly :: ScopDetect/mod_ref_read_pointer.ll
- Polly :: ScopInfo/max-loop-depth.ll
- Polly :: ScopInfo/mod_ref_access_pointee_arguments.ll
- Polly :: ScopInfo/mod_ref_read_pointee_arguments.ll
- Polly :: ScopInfo/mod_ref_read_pointer.ll
- Polly :: ScopInfo/mod_ref_read_pointers.ll
The test cases failed after removing the heuristic.
Differential Revision: https://reviews.llvm.org/D45274
Contributed-by: Lorenzo Chelini <l.chelini@icloud.com>
llvm-svn: 329548
The warning is:
isl_union_map.c(2041): warning C4221: nonstandard extension used: 'filter_user': cannot be initialized using address of automatic variable 'data'
for the following code (and others)
struct isl_un_op_drop_user_data data = { &isl_set_is_wrapping };
struct isl_un_op_control control = {
.filter = &un_op_filter_drop_user,
.filter_user = &data,
.fn_map = &isl_set_wrapped_domain_map,
};
llvm-svn: 329328
The
}; // namespace polly
comment was put at the closing brace of the FunctionToScopPassAdaptor class.
Since no namespace ends here, the comment is misplaced.
Reported-by: Lukas Böhm <lukas.boehm93@gmail.com>
llvm-svn: 329302
Summary:
When checking the parallelism of a scheduling dimension, we first check if excluding reduction dependences the loop is parallel or not.
If the loop is not parallel, then we need to return the minimal dependence distance of all data dependences, including the previously subtracted reduction dependences.
Reviewers: grosser, Meinersbur, efriedma, eli.friedman, jdoerfert, bollu
Reviewed By: Meinersbur
Subscribers: llvm-commits, pollydev
Tags: #polly
Differential Revision: https://reviews.llvm.org/D45236
llvm-svn: 329214
This gets very expensive for basic blocks which don't have a name: it
calls printAsOperand, which numbers the entire module. We don't
normally need the name anyway, though; it's only used for debug dumps,
so don't compute it by default.
Differential Revision: https://reviews.llvm.org/D44946
llvm-svn: 328666
Summary:
r327219 added wrappers to std::sort which randomly shuffle the container before sorting.
This will help in uncovering non-determinism caused due to undefined sorting
order of objects having the same key.
To make use of that infrastructure we need to invoke llvm::sort instead of std::sort.
Reviewers: grosser, efriedma, jdoerfert, bollu, sebpop
Reviewed By: sebpop
Subscribers: sebpop, mehdi_amini, llvm-commits, pollydev
Tags: #polly
Differential Revision: https://reviews.llvm.org/D44361
llvm-svn: 327361
Piecewise affine expressions have directly corresponding mathematical
operators. Introduce these operators as overloads as this makes writing
code with isl::pw_aff expressions more directly readable.
We can now write:
A = B + C instead of A = B.add(C)
llvm-svn: 327216
Summary:
When building polly as part of the monorepo (actually, as part of any setup
using LLVM_ENABLE_PROJECTS), the LLVMPolly library used in the lit tests ends
up in a different directory in the build tree than in an in-tree build
Reviewers: Meinersbur, grosser, bollu
Reviewed By: Meinersbur
Subscribers: mgorny, bollu, pollydev, llvm-commits
Differential Revision: https://reviews.llvm.org/D44078
llvm-svn: 326702
isl does not guarantee that set dimension ids will be preserved, so using them
to carry information is not a good idea. Furthermore, the loop information can
be derived without problem from the statement itself. As this even requires
less code than propagating loop information on set dimension ids, starting from
this commit we just derive the loop information in collectSurroundingLoops
directly from the IR.
Interestingly this also results in a couple of isl sets to take a simpler
representation.
llvm-svn: 326664
During codegen, Polly attempts to clear all loops from ScalarEvolution
and LoopInfo, and it does so one block at a time. This causes undefined
behaviour, since this way a loop header might be removed from a loop
before the entire loop is erased, causing ScalarEvolution to run into an
error.
Instead, just delete the entire loop atomically. This fixes currently
failing testcases.
llvm-svn: 326643
Also un-revert (isl_pw_*_alloc: add missing check for compatible spaces, Wed Sep
6 12:18:04 2017 +0200).
This patch is a proposed fix to avoid asserts due to stricter space checking
within isl, which resulted in failures when converting a schedule tree to
a schedule map.
llvm-svn: 326073
As part of this cleanup a couple of unnecessary isl::manage(obj.copy()) pattern
are eliminated as well.
We checked for all potential cleanups by scanning for:
"grep -R isl::manage\( lib/ | grep copy"
llvm-svn: 325558
This update:
- Removes several deprecated functions (e.g., isl_band).
- Improves the pretty-printing of sets by detecting modulos and "false"
equalities.
- Minor improvements to coalescing and increased robustness of the isl
scheduler.
This update does not yet include isl commit isl-0.18-90-gd00cb45
(isl_pw_*_alloc: add missing check for compatible spaces, Wed Sep 6 12:18:04
2017 +0200), as this additional check is too tight and unfortunately causes
two test case failures in Polly. A patch has been submitted to isl and will be
included in the next isl update for Polly.
llvm-svn: 325557
Two or more PHIs mutually using each other directly or indirectly as
incoming value could cause that a PHI WRITE be added before the PHI READ
(i.e. it overwrites the current incoming value with the next incoming
value before it being read).
Fix by ensuring that the PHI WRITE and PHI READ are in the same statement.
This should fix the miscompile of SingleSource/Benchmark/Misc/whetstone
from the test-suite.
llvm-svn: 324934
Splitting basic blocks into multiple statements if there are now
additional scalar dependencies gives more freedom to the scheduler, but
more statements also means higher compile-time complexity. Switch to
finer statement granularity, the additional compile time should be
limited by the number of operations quota.
The regression tests are written for the -polly-stmt-granularity=bb
setting, therefore we add that flag to those tests that break with the
new default. Some of the tests only fail because the statements are
named differently due to a basic block resulting in multiple statements,
but which are removed during simplification of statements without
side-effects. Previous commits tried to reduce this effect, but it is
not completely avoidable.
Differential Revision: https://reviews.llvm.org/D42151
llvm-svn: 324169
Do not add a "_last" suffix to the statement name if there is no (other)
main statement for a basic block. In other words, it becomes the main
statement itself. This further reduces the statement naming difference
between -polly-stmt-granularity=bb and
-polly-stmt-granularity=scalar-indep.
llvm-svn: 324168
Summary:
This change is step four in the series of changes to remove alignment argument from
memcpy/memmove/memset in favour of alignment attributes. Steps:
Step 1) Remove alignment parameter and create alignment parameter attributes for
memcpy/memmove/memset. ( rL322965, rC322964, rL322963 )
Step 2) Expand the IRBuilder API to allow creation of memcpy/memmove with differing
source and dest alignments. ( rL323597 )
Step 3) Update Clang to use the new IRBuilder API. ( rC323617 )
Step 4) Update Polly to use the new IRBuilder API.
Step 5) Update LLVM passes that create memcpy/memmove calls to use the new IRBuilder API,
and those that use use MemIntrinsicInst::[get|set]Alignment() to use [get|set]DestAlignment()
and [get|set]SourceAlignment() instead.
Step 6) Remove the single-alignment IRBuilder API for memcpy/memmove, and the
MemIntrinsicInst::[get|set]Alignment() methods.
Reference
http://lists.llvm.org/pipermail/llvm-dev/2015-August/089384.htmlhttp://lists.llvm.org/pipermail/llvm-commits/Week-of-Mon-20151109/312083.html
Reviewers: jdoerfert, grosser, bollu
Subscribers: llvm-commits
Differential Revision: https://reviews.llvm.org/D41677
llvm-svn: 323618
Theoretically, a PHI write can be added to any statement that represents
the incoming basic block. We previously always chose the last because
the incoming value's definition is guaranteed to be defined.
With this patch the PHI write is added to the statement that defines the
incoming value. It avoids the requirement for a scalar dependency between
the defining statement and the statement containing the write. As such the
logic for -polly-stmt-granularity=scalar-indep that ensures that there is
such scalar dependencies can be removed.
Differential Revision: https://reviews.llvm.org/D42147
llvm-svn: 323284
VirtualUse::create is only called for MemoryKind::Value, but its
consistency nonetheless checked in verifyUses(). PHI uses are always
inter-stmt dependencies, which was not considered by the constructor
method. The virtual and non-virtual execution paths were the same, such
that verifyUses did not encounter any inconsistencies.
llvm-svn: 323283
except Darwin and Windows. This prevents inserting an environment
variable with an empty name (which is illegal and leads to a Python
exception) on any of the BSDs.
llvm-svn: 323041
Summary:
Upstream LLVM is changing the the prototypes of the @llvm.memcpy/memmove/memset
intrinsics. This change updates the polly tests for this change.
The @llvm.memcpy/memmove/memset intrinsics currently have an explicit argument
which is required to be a constant integer. It represents the alignment of the
dest (and source), and so must be the minimum of the actual alignment of the
two.
This change removes the alignment argument in favour of placing the alignment
attribute on the source and destination pointers of the memory intrinsic call.
For example, code which used to read:
call void @llvm.memcpy.p0i8.p0i8.i32(i8* %dest, i8* %src, i32 100, i32 4, i1 false)
will now read
call void @llvm.memcpy.p0i8.p0i8.i32(i8* align 4 %dest, i8* align 4 %src, i32 100, i1 false)
At this time the source and destination alignments must be the same (Step 1).
Step 2 of the change, to be landed shortly, will relax that contraint and allow
the source and destination to have different alignments.
llvm-svn: 322963
The goal is to have -polly-stmt-granularity=bb and
-polly-stmt-granularity=scalar-indep to have the same names if there is
just one statement per basic block.
This fixes a fluke when Polybench's jacobi-2d is optimized differently
depending on the -polly-stmt-granularity option, although both options
create the same SCoP, just with different statement names.
The new naming scheme is:
With -polly-use-llvm-names=0:
Stmt<BBIdx as decimal><Idx within BB as letter>
With -polly-use-llvm-names=1:
Stmt_BBName_<Idx within BB as letter>
The <Idx within BB> suffix is omitted for the main statement of a BB. The
main statement is either the one containing the first store or call
(those cannot be removed by the simplifyer), or if there is no such
instruction, the first. If after simplification there is just a single
statement left, it should be the main statement and have the same names as
with -polly-stmt-granularity=bb.
Differential Revision: https://reviews.llvm.org/D42136
llvm-svn: 322852
This will give control of the statement's name to the caller.
Required to give -polly-stmt-granularity=scalar-indep more control
over the name of the generated statement in a follow-up commit.
llvm-svn: 322851
isl_val_get_num_si crashes on overflow, so don't use it on arbitrary
integers.
Testcase only crashes on platforms where long is 32 bits because of the
signature of isl_val_get_num_si; not sure if it's possible to write a
testcase which crashes if long is 64 bits.
There are a few other places in polly which use isl_val_get_num_si;
they probably need to be fixed as well. I don't think polly uses any
of the other "long" isl APIs in an unsafe manner.
Differential Revision: https://reviews.llvm.org/D42129
llvm-svn: 322766
Print same or similar structure elements together. Previously, the
value could take more importance that the space structure if visited
first in the space nest tree.
Before:
{
Left[0] -> Right[i]: i >= 0;
Left[1] -> AnotherRight[i];
Left[2] -> Right[-1]
}
After:
{
Left[0] -> Right[i]: i >= 0;
Left[2] -> Right[-1];
Left[1] -> AnotherRight[i]
}
llvm-svn: 322581
CMake insists that for each target, one uses only the non-keyword
version of target_link_library
target_link_library(mytarget lib)
or the one with PUBLIC/PRIVATE/INTERFACE keyword:
target_link_library(mytarget PUBLIC lib)
Otherwise, CMake fails with the error message:
The keyword signature for target_link_libraries has already been used with
the target "mytarget". All uses of target_link_libraries with a target
must be either all-keyword or all-plain.
Change all occurances of target_link_library to the newer keyworded
version to avoid such errors. Some already have been changed in r319840,
but might not be sufficient for all build configurations to build
the doxygen manual.
Reported-by: Tanya Lattner <tanyalattner@llvm.org>
llvm-svn: 322376
Memory transfer instructions take two pointers. It is not defined to
which of those a noalias annotation applies. To ensure correctness,
do not add noalias annotations to memcpy/memmove instructions anymore.
The caused a miscompile with test-suite's MultiSource/Applications/obsequi.
Since r321138, the MemCpyOpt pass would remove memcpy/memmove calls if
known to copy uninitialized memory. In that case, it was initialized
by another memcpy, but the annotation for the target pointer said
it would not alias. The annotation was actually meant for the source
pointer, which was was an alloca and could not alias with the target
pointer.
llvm-svn: 321371
If an out-of-quota error occurred, the last error would be
isl_error_quota unless a different error occured. We typically check
whether the max-operations occured by comparing to that error value
after leaving the quota guard. This would check whether there ever
was a quota-error, not just in the last quota guards.
The observable bug occurred if the max-operations limit was reached in
DeLICM, and if -polly-dependences-computout=0, DependenceInfo would
think that the quota for computing dependencies was the reason,
i.e., fail the operation even if the calculation itself was successful.
Fix by reseting the last error to isl_error_none when entering a
quota guard, signaling that no quota error occured unless in the
guard's scope.
llvm-svn: 321329
We currently use target_link_libraries without an explicit scope
specifier (INTERFACE, PRIVATE or PUBLIC) when linking executables.
Dependencies added in this way apply to both the target and its
dependencies, i.e. they become part of the executable's link interface
and are transitive.
Transitive dependencies generally don't make sense for executables,
since you wouldn't normally be linking against an executable. This also
causes issues for generating install export files when using
LLVM_DISTRIBUTION_COMPONENTS. For example, clang has a lot of LLVM
library dependencies, which are currently added as interface
dependencies. If clang is in the distribution components but the LLVM
libraries it depends on aren't (which is a perfectly legitimate use case
if the LLVM libraries are being built static and there are therefore no
run-time dependencies on them), CMake will complain about the LLVM
libraries not being in export set when attempting to generate the
install export file for clang. This is reasonable behavior on CMake's
part, and the right thing is for LLVM's build system to explicitly use
PRIVATE dependencies for executables.
Unfortunately, CMake doesn't allow you to mix and match the keyword and
non-keyword target_link_libraries signatures for a single target; i.e.,
if a single call to target_link_libraries for a particular target uses
one of the INTERFACE, PRIVATE, or PUBLIC keywords, all other calls must
also be updated to use those keywords. This means we must do this change
in a single shot. I also fully expect to have missed some instances; I
tested by enabling all the projects in the monorepo (except dragonegg),
and configuring both with and without shared libraries, on both Darwin
and Linux, but I'm planning to rely on the buildbots for other
configurations (since it should be pretty easy to fix those).
Even after this change, we still have a lot of target_link_libraries
calls that don't specify a scope keyword, mostly for shared libraries.
I'm thinking about addressing those in a follow-up, but that's a
separate change IMO.
Differential Revision: https://reviews.llvm.org/D40823
llvm-svn: 319840
Summary:
This can be seen as a follow-up on my previous differential [D33411](https://reviews.llvm.org/D33411).
We received a bug report where this error was triggered. I have tried my best to recreate the issue in a minimal lit testcase which is also part of this differential.
I only handle return instructions as predecessors to a virtual TLR-exit right now. From inspecting the codebase, it seems `unreachable` instructions may also be of interest here. If requested, I can extend my patches to consider them as well. I would also apply this on `ScopHelper.cpp::isErrorBlock` (see D33411), of course.
Reviewers: philip.pfaffe, bollu
Reviewed By: bollu
Subscribers: Meinersbur, pollydev, llvm-commits
Tags: #polly
Differential Revision: https://reviews.llvm.org/D40492
llvm-svn: 319431
Summary: We want to automatically copy the appropriate mailing list
for review requests to the polly repository.
For context, see the proposal and discussion here:
http://lists.llvm.org/pipermail/cfe-dev/2017-November/056032.html
Similar to D40179, I set up a new Diffusion repository with callsign
"PLO" for polly:
https://reviews.llvm.org/source/polly/
This explicitly updates polly's .arcconfig to point to the new C
repository in Diffusion, which will let us use Herald rule H270.
llvm-svn: 319056
Isl does not allow generating isl_ast_expr from an isl_pw_aff that has an
empty domain (i.e. has no pieces). We already detected the case if the
isl_pw_aff comes with an empty domain.
isl_ast_build also considers the domain empty if it is disjoint with the
parameter context (e.g. parameters values that we exclude by runtime
versioning).
Intersect the access relation domain with the parameter context to
also detect such practically empty access domains. The effective
pointer used in the generated code is unimportand because it will never
be executed.
This fixes llvm.org/PR35362
llvm-svn: 318806
Summary:
Most changes are mechanical, but in one place I changed the program semantics
by fixing a likely bug:
In `Scop::hasFeasibleRuntimeContext()`, I'm now explicitely handling the
error-case. Before, when the call to `addNonEmptyDomainConstraints()`
returned a null set, this (probably) accidentally worked because
isl_bool_error converts to true. I'm checking for nullptr now.
Reviewers: grosser, Meinersbur, bollu
Reviewed By: Meinersbur
Subscribers: nemanjai, kbarton, pollydev, llvm-commits
Differential Revision: https://reviews.llvm.org/D39971
llvm-svn: 318632
Summary:
There is a potential use-after-free bug in Scop::buildSchedule(Region *,
LoopStackTy &, LoopInfo &). Before, we took a reference to LoopStack.back()
which is a use after free, since back is popped off further below. This didn't
crash before by pure chance, since LoopStack is actually a vector, and the
memory isn't freed upon pop. I turned this into an iterator-based algorithm.
Reviewers: grosser, bollu, Meinersbur
Reviewed By: Meinersbur
Subscribers: llvm-commits, pollydev
Differential Revision: https://reviews.llvm.org/D39979
llvm-svn: 318415
Put the analysis part of reloadKnownContent under an isl
max-operations quota scope, as has already been done for
forwardKnownLoad.
This should fix the aosp timeout of "GrTestUtils.cpp".
llvm-svn: 317495
Represent PHIs by their incoming values instead of an opaque value of
themselves. This allows ForwardOpTree to "look through" the PHIs and
forward the incoming values since forwardings PHIs is currently not
supported.
This is particularly useful to cope with PHIs inserted by GVN LoadPRE.
The incoming values all resolve to a load from a single array element
which then can be forwarded.
It should in theory also reduce spurious conflicts in value mapping
(DeLICM), but I have not yet found a profitable case yet, so it is
not included here.
To avoid transitive closure and potentially necessary overapproximations
of those, PHIs that may reference themselves are excluded from
normalization and keep their opaque self-representation.
Differential Revision: https://reviews.llvm.org/D39333
llvm-svn: 317008
ForwardOpTree may already transform a scalar access to an array
accesses. The access remains implicit (isOriginalScalarKind(), meaning
that the access is always executed at the begin/end of a statement), but
targets an array (isLatestArrayKind(), which is unrelated to whether the
execution is implicit/explicit).
Fix by properly using isOriginalXXX() to determine execution order.
This fixes the buildbots on MultiSource/Benchmarks/DOE-ProxyApps-C/miniGMG.
llvm-svn: 316995
When collecting base pointers that need to be made available in parallel
subfunctions, use the base pointer associated with the latest
ScopArrayInfo, instead of the original one.
llvm-svn: 316983
Summary:
When GPUNodeBuilder creates loops inside the kernel, it dispatches to
IslNodeBuilder. This however is surprisingly dangerous, since it accesses the
AST Node's user through the wrong type. This patch fixes this problem by
overriding createFor correctly.
This fixes PR35010.
Reviewers: grosser, bollu, Meinersbur
Reviewed By: Meinersbur
Subscribers: Meinersbur, nemanjai, pollydev, llvm-commits, kbarton
Differential Revision: https://reviews.llvm.org/D39364
llvm-svn: 316872
Add missing %loadPolly directive to support out of tree builds. One of
the changes is somewhat bigger, because the directive turns on LLVM
names, and the testcase deosn't use those.
llvm-svn: 316870
For scalar accesses, change the access target to an array element that
is known to contain the same value.
This may become an alternative to forwardKnownLoad which creates new
loads (and therefore closer to forwarding speculatives). Reloading does
not require the known value originating from a load, but can be a store
as well.
Differential Revision: https://reviews.llvm.org/D39325
llvm-svn: 316766
Previously we marked scalars based on the original access function. However,
when a scalar read access is redirected, the original definition
(or incoming values of a PHI) is not used anymore, and can be deleted
(unless referenced by use that has not been redirected).
llvm-svn: 316660
Add check and skip when the store used to determine the target accesses
multiple array elements. Only a single array location should for
mapping the scalar. Having multiple creates problems when deciding which
element to load from. While MemoryAccess::getAddressFunction() should
select just one of them, other problems arise in code that assumes
that there is just one target element per statement instance.
This fixes llvm.org/PR34989
This also reverts r313902 which fixed llvm.org/PR34485 also caused by
a non-functional target array element. This patch avoids the situation
to occur in the first place.
llvm-svn: 316432
After rL315683 (improve SCEV to calculate max BETakenCount when end
bound of loop is variant and loop is of form {Start,+1, Stride} LT End)
this test in polly started failing.
However, as discussed in https://reviews.llvm.org/rL315683,
this polly test is not a loops bound test and the MaxBECount calculated by
SCEV looks correct. The max BECount is the value calculated even when the end
bound of loop is invariant.
As discussed with Tobias offline, I'm marking this as an XFAIL, until he
gets a chance to update the testcase, so the build bot goes to green.
llvm-svn: 315912
The option splits BasicBlocks into minimal statements such that no
additional scalar dependencies are introduced.
The algorithm is based on a union-find structure, and unites sets if
putting them into separate statements would introduce a scalar
dependencies. As a consequence, instructions may be split into separate
statements such their relative order is different than the statements
they are in. This is accounted for instructions whose relative order
matters (e.g. memory accesses).
The algorithm is generic in that heuristic changes can be made
relatively easily. We might relax the order requirement for read-reads
or accesses to different base pointers. Forwardable instructions can be
made to not cause a join.
This implementation gives us a speed-up of 82% in SPEC 2006 456.hmmer
benchmark by allowing loop-distribution in a hot loop such that one of
the loops can be vectorized.
Differential Revision: https://reviews.llvm.org/D38403
llvm-svn: 314983
The option is introduced with only one possible value
-polly-stmt-granularity=bb which represents the current behaviour, which
is outlined into the new function buildSequentialBlockStmts().
More options will be added in future commits.
llvm-svn: 314900
We make sure that the final reload of an invariant scalar memory access uses the
same stack slot into which the invariant memory access was stored originally.
Earlier, this was broken as we introduce a new stack slot aside of the preload
stack slot, which remained uninitialized and caused our escaping loads to
contain garbage. This happened due to us clearing the pre-populated values
in EscapeMap after kernel code generation. We address this issue by preserving
the original host values and restoring them after kernel code generation.
EscapeMap is not expected to be used during kernel code generation, hence we
clear it during kernel generation to make sure that any unintended uses are
noticed.
llvm-svn: 314894
This test XFAILs two test that start to fail when verifying DT's
DFS numbers, as per Tobias' suggestion.
Related VerifyDFSNumbers patch: D38331.
llvm-svn: 314800
Iterate over statement instructions instead over basic block
instructions when creating MemoryAccesses. It allows making the creation
of MemoryAccesses independent of how the basic blocks are split into
multiple ScopStmts.
llvm-svn: 314665
Create the MemoryAccesses of invariant loads separately and before
all other MemoryAccesses.
Invariant loads are classified as synthesizable and therefore are not
contained in any statement. When iterating over all instructions of all
statements, the invariant loads are consequently not processed and
iterating over them separately becomes necessary.
This patch can change the order in which MemoryAccesses are created, but
otherwise has no functional change.
Some temporary code is introduced to ensure correctness, but will be
removed in the next commit.
llvm-svn: 314664
Instructions that compute escaping values might be synthesizable and
therefore not contained in any ScopStmt. When buildAccessFunctions is
changed to only iterate over the instruction list of statement,
"free" instructions still need to be written. We do this after the
main MemoryAccesses have been created.
This can change the order in which MemoryAccesses are created, but has
otherwise no functional change.
llvm-svn: 314663
Decouple handling of exit block PHIs and other MemoryAccesses. Exit PHIs
only need the PHI handling part of buildAccessFunctions but requires
code for skipping them in while creating other MemoryAcesses.
This change will make it easier to modify how statement MemoryAccesses
are created without considering the exit block special case.
llvm-svn: 314662
Loads before the SCoP are always invariant within the SCoP and
therefore are no "required invariant loads". An assertion failes in
ScopBuilder when it finds such an invariant load.
Fix by not adding such loads to the required invariant load list. This
likely will cause the region to be not considered a valid SCoP.
We may want to unconditionally accept instructions defined before
the region as valid invariant conditions instead of rejecting them.
This fixes a compilation crash of SPEC CPU2006 453.povray's
render.cpp.
llvm-svn: 314636
This matches the behavior we already have in lib/Codegen/CodeGeneration.cpp and
makes sure that we fall back to the original code. It seems when invariant load
hoisting was introduced to the GPGPU backend we missed to reset the RTC flag,
such that kernels where invariant load hoisting failed executed the 'optimized'
SCoP, which however is set to a simple 'unreachable'. Unsurprisingly, this
results in hard to debug issues that are a lot of fun to debug.
llvm-svn: 314624
These functions print a multi-line and sorted representation of unions
of polyhedra. Each polyhedron (basic_{ast/map}) has its own line.
First sort key is the polyhedron's hierachical space structure.
Secondary sort key is the lower bound of the polyhedron, which should
ensure that the polyhedral are printed in approximately ascending order.
Example output of dumpPw():
[p_0, p_1, p_2] -> {
Stmt0[0] -> [0, 0];
Stmt0[i0] -> [i0, 0] : 0 < i0 <= 5 - p_2;
Stmt1[0] -> [0, 2] : p_1 = 1 and p_0 = -1;
Stmt2[0] -> [0, 1] : p_1 >= 3 + p_0;
Stmt3[0] -> [0, 3];
}
In contrast dumpExpanded() prints each point in the sets, unless there
is an unbounded dimension that cannot be expandend.
This is useful for reduced test cases where the loop counts are set to
some constant to understand a bug.
Example output of dumpExpanded(
{ [MemRef_A[i0] -> [i1]] : (exists (e0 = floor((1 + i1)/3): i0 = 1 and
3e0 <= i1 and 3e0 >= -1 + i1 and i1 >= 15 and i1 <= 25)) or (exists (e0
= floor((i1)/3): i0 = 0 and 3e0 < i1 and 3e0 >= -2 + i1 and i1 > 0 and
i1 <= 11)) }):
{
[MemRef_A[0] ->[1]];
[MemRef_A[0] ->[2]];
[MemRef_A[0] ->[4]];
[MemRef_A[0] ->[5]];
[MemRef_A[0] ->[7]];
[MemRef_A[0] ->[8]];
[MemRef_A[0] ->[10]];
[MemRef_A[0] ->[11]];
[MemRef_A[1] ->[15]];
[MemRef_A[1] ->[16]];
[MemRef_A[1] ->[18]];
[MemRef_A[1] ->[19]];
[MemRef_A[1] ->[21]];
[MemRef_A[1] ->[22]];
[MemRef_A[1] ->[24]];
[MemRef_A[1] ->[25]]
}
Differential Revision: https://reviews.llvm.org/D38349
llvm-svn: 314525
Summary:
Add a document which describes:
- GEMM performance comparison.
- An experiment that measures the compile time impact
of enabling Polly when compiling LLVM+Clang+Polly.
Contributed-by: Theodoros Theodoridis<theodoros.theodoridis@inf.ethz.ch>
Differential Revision: https://reviews.llvm.org/D38330
llvm-svn: 314419
In order for debuggers to be able to call an inline method, it must have
been instantiated somewhere. The dump() methods are usually not used, so
add an instantiation in debug builds.
This allows to call .dump() on any isl++ object from the gcc/gdb and
Visual Studio debugger in debug builds with assertions enabled.
In optimized builds, even with assertions enabled, the dump() methods
are also inlined in GICHelper.cpp, so no externally visible symbols
will be available either.
Differential Revision: https://reviews.llvm.org/D38198
llvm-svn: 314395
In case a PHI node follows an error block we can assume that the incoming value
can only come from the node that is not an error block. As a result, conditions
that seemed non-affine before are now in fact affine.
This is a recommit of r312663 after fixing
test/Isl/CodeGen/phi_after_error_block_outside_of_scop.ll
llvm-svn: 314075
Such RTCs may introduce integer wrapping intrinsics with more than 64 bit,
which are translated to library calls on AOSP that are not part of the
runtime and will consequently cause linker errors.
Thanks to Eli Friedman for reporting this issue and reducing the test case.
llvm-svn: 314065
Remove an assertion that tests the injectivity of the
PHIRead -> PHIWrite relation. That is, allow a single PHI write to be
used by multiple PHI reads. This may happen due to some statements
containing the PHI write not having the statement instances that would
overwrite the previous incoming value due to (assumed/invalid) contexts.
This result in that PHI write is mapped to multiple targets which is not
supported. Codegen will select one one of the targets using
getAddressFunction(). However, the runtime check should protect us from
this case ever being executed.
We therefore allow injective PHI relations. Additional calculations to
detect/santitize this case would probably not be worth the compuational
effort.
This fixes llvm.org/PR34485
llvm-svn: 313902
Before this patch, ScopInfo::getValueDef(SAI) used
getStmtFor(Instruction*) to find the MemoryAccess that writes a
MemoryKind::Value. In cases where the value is synthesizable within the
statement that defines, the instruction is not added to the statement's
instruction list, which means getStmtFor() won't return anything.
If the synthesiable instruction is not synthesiable in a different
statement (due to being defined in a loop that and ScalarEvolution
cannot derive its escape value), we still need a MemoryKind::Value
and a write to it that makes it available in the other statements.
Introduce a separate map for this purpose.
This fixes MultiSource/Benchmarks/MallocBench/cfrac where
-polly-simplify could not find the writing MemoryAccess for a use. The
write was not marked as required and consequently was removed.
Because this could in principle happen as well for PHI scalars,
add such a map for PHI reads as well.
llvm-svn: 313881
Since -polly-codegen reports itself to preserve DependenceInfo and IslAstInfo,
we might get those analysis that were computed by a different ScopInfo for a
different Scop structure. This would be unfortunate because DependenceInfo and
IslAstInfo hold references to resources allocated by
ScopInfo/ScopBuilder/Scop (e.g. isl_id). If -polly-codegen and
DependenceInfo/IslAstInfo do not agree on which Scop to use, unpredictable
things can happen.
When the ScopInfo/Scop object is freed, there is a high probability that the
new ScopInfo/Scop object will be created at the same heap position with the
same address. Comparing whether the Scop or ScopInfo address is the expected
therefore is unreliable.
Instead, we compare the address of the isl_ctx object. Both, DependenceInfo
and IslAstInfo must hold a reference to the isl_ctx object to ensure it is
not freed before the destruction of those analyses which might happen after
the destruction of the Scop/ScopInfo they refer to. Hence, the isl_ctx
will not be freed and its address not reused as long there is a
DependenceInfo or IslAstInfo around.
This fixes llvm.org/PR34441
llvm-svn: 313842
Fix walking over the schedule tree to collect its properties
(Number of permutable bands etc.).
Also add regression tests for these statistics.
llvm-svn: 313750
Computing the reaching definition in forwardTree() can take a long time
if the coefficients are large. When the forwarding is
carried-out (doIt==true), forwardTree() must execute entirely or not at
all to get a consistent output, which means we cannot just allow
out-of-quota errors to happen in the middle of the processing.
We introduce the class IslQuotaScope which allows to opt-in code that is
conformant and has been tested with out-of-quota events. In case of
ForwardOpTree, out-of-quota is allowed during the operand tree
examination, but not during the transformation. The same forwardTree()
recursion is used for examination and execution, meaning that the
reaching definition has already been computed in the examination tree
walk and cached for reuse in the transformation tree walk.
This should fix the time-out of grtestutils.ll of the asop buildbot. If
the compilation still takes too long, we can reduce the max-operations
allows for -polly-optree.
Differential Revision: https://reviews.llvm.org/D37984
llvm-svn: 313690
cl::opt<unsigned long> is not specialized and hence the option
-polly-optree-max-ops impossible to use.
Replace by supported option cl::opt<unsigned>.
Also check for an error state when computing the written value, which
happens when the quota runs out.
llvm-svn: 313546
In r301670 IR verification was disabled. Since then, CodeGen writing
malformed IR would only be noticed by unpredictable behavior in
follow-up passes (e.g. segfaults, infinite loops) or IR verification in
the backend assert builds.
Re-enable -polly-codegen-verify at for the regression tests to ensure
that malformed IR is detected where Polly generated malformed IR in the
past and changes in CodeGen are at least partially covered by
check-polly
(otherwise malformed IR may only get noticed when the buildbots run the
test-suite).
Differential Revision: https://reviews.llvm.org/D37969
llvm-svn: 313527
This is a resubmission of r313270. It broke standalone builds of
compiler-rt because we were not correctly generating the llvm-lit
script in the standalone build directory.
The fixes incorporated here attempt to find llvm/utils/llvm-lit
from the source tree returned by llvm-config. If present, it
will generate llvm-lit into the output directory. Regardless,
the user can specify -DLLVM_EXTERNAL_LIT to point to a specific
lit.py on their file system. This supports the use case of
someone installing lit via a package manager. If it cannot find
a source tree, and -DLLVM_EXTERNAL_LIT is either unspecified or
invalid, then we print a warning that tests will not be able
to run.
Differential Revision: https://reviews.llvm.org/D37756
llvm-svn: 313407
This patch is still breaking several multi-stage compiler-rt bots.
I already know what the fix is, but I want to get the bots green
for now and then try re-applying in the morning.
llvm-svn: 313335
This patch simplifies LLVM's lit infrastructure by enforcing an ordering
that a site config is always run before a source-tree config.
A significant amount of the complexity from lit config files arises from
the fact that inside of a source-tree config file, we don't yet know if
the site config has been run. However it is *always* required to run
a site config first, because it passes various variables down through
CMake that the main config depends on. As a result, every config
file has to do a bunch of magic to try to reverse-engineer the location
of the site config file if they detect (heuristically) that the site
config file has not yet been run.
This patch solves the problem by emitting a mapping from source tree
config file to binary tree site config file in llvm-lit.py. Then, during
discovery when we find a config file, we check to see if we have a
target mapping for it, and if so we use that instead.
This mechanism is generic enough that it does not affect external users
of lit. They will just not have a config mapping defined, and everything
will work as normal.
On the other hand, for us it allows us to make many simplifications:
* We are guaranteed that a site config will be executed first
* Inside of a main config, we no longer have to assume that attributes
might not be present and use getattr everywhere.
* We no longer have to pass parameters such as --param llvm_site_config=<path>
on the command line.
* It is future-proof, meaning you don't have to edit llvm-lit.in to add
support for new projects.
* All of the duplicated logic of trying various fallback mechanisms of
finding a site config from the main config are now gone.
One potentially noteworthy thing that was required to implement this
change is that whereas the ninja check targets previously used the first
method to spawn lit, they now use the second. In particular, you can no
longer run lit.py against the source tree while specifying the various
`foo_site_config=<path>` parameters. Instead, you need to run
llvm-lit.py.
Differential Revision: https://reviews.llvm.org/D37756
llvm-svn: 313270
The remaining parts produced by the full partial tile isolation can contain
hot spots that are worth to be optimized. Currently, we rely on the simple
loop unrolling pass, LiCM and the SLP vectorizer to optimize such parts.
However, the approach can suffer from the lack of the information about
aliasing that Polly provides using additional alias metadata or/and the lack
of the information required by simple loop unrolling pass.
This patch is the first step to optimize the remaining parts. To do it, we
unroll and separate them. In case of, for instance, Intel Kaby Lake, it helps
to increase the performance of the generated code from 39.87 GFlop/s to
49.23 GFlop/s.
The next possible step is to avoid unrolling performed by Polly in case of
isolated and remaining parts and rely only on simple loop unrolling pass and
the Loop vectorizer.
Reviewed-by: Tobias Grosser <tobias@grosser.es>
Differential Revision: https://reviews.llvm.org/D37692
llvm-svn: 312929
Update CodegenCleanup using the function-level passes added by
populatePassManager that run between EP_EarlyAsPossible and
EP_VectorizerStart in -O3.
The changes in particular are:
- Added pass create arguments, e.g. ExpensiveCombines for InstCombine.
- Remove reroll pass. The option -reroll-loops is disabled by default.
- Add passes run with UnitAtATime, which is the default.
- Add instances of LibCallsShrinkWrap, TailCallElimination, SCCP
(sparse conditional constant propagation), Float2Int
that did not run before.
- Add instances of GVN as in the default pipeline.
Notes:
- GVNHoist, GVNSink, NewGVN are still disabled in the -O3 pipeline.
- The optimization level and other optimization parameters are not
accessible outside of PassManagerBuilder, hence we cannot add passes
depending on these.
Differential Revision: https://reviews.llvm.org/D37571
llvm-svn: 312875
The type of NewValue might change due to ScalarEvolution
looking though bitcasts. The synthesized NewValue therefore
becomes the type before the bitcast.
llvm-svn: 312718
This reverts commit
r312410 - [ScopDetect/Info] Look through PHIs that follow an error block
The commit caused generation of invalid IR due to accessing a parameter
that does not dominate the SCoP.
llvm-svn: 312663
Up to now ZoneAlgo considered array elements access by something else
than a LoadInst or StoreInst as not analyzable. This patch removes that
restriction by using the unknown ValInst to describe the written
content, repectively the element type's null value in case of memset.
Differential Revision: https://reviews.llvm.org/D37362
llvm-svn: 312630
Tobias Grosser