Right now the Attributor defaults to 32 fixed point iterations unless it is set
explicitly by a command line flag. This patch allows this to be configured when
the attributor instance is created. The maximum is then increased in OpenMPOpt
if the target is a kernel. This is because the globalization analysis can result
in larger iteration counts due to many dependent instances running at once.
Depends on D102444
Reviewed By: jdoerfert
Differential Revision: https://reviews.llvm.org/D104416
Summary:
This patch adds support for the Attributor to emit remarks on behalf of some
other pass. The attributor can now optionally take a callback function that
returns an OptimizationRemarkEmitter object when given a Function pointer. If
this is availible then a remark will be emitted for the corresponding pass
name.
Depends on D102197
Reviewed By: sstefan1 thegameg
Differential Revision: https://reviews.llvm.org/D102444
Summary:
The changes to globalization introduced in D97680 introduce a large amount of overhead by default. The old globalization method would always ignore globalization code if executing in SPMD mode. This wasn't strictly correct as data sharing is still possible in SPMD mode. The new interface is correct but introduces globalization code even when unnecessary. This optimization will use the existing HeapToStack transformation in the attributor to allow for unneeded globalization to be replaced with thread-private stack memory. This is done using the newly introduced library instances for the RTL functions added in D102087.
Depends on D97818
Reviewed By: jdoerfert
Differential Revision: https://reviews.llvm.org/D102197
Summary:
Currently the attributor needs to give up if a function has external linkage.
This means that the optimization introduced in D97818 will only apply to static
functions. This change uses the Attributor to internalize OpenMP device
routines by making a copy of each function with private linkage and replacing
the uses in the module with it. This allows for the optimization to be applied
to any regular function.
Reviewed By: jdoerfert
Differential Revision: https://reviews.llvm.org/D102824
Summary:
The changes introduced in D97680 create a simpler interface to code that needs
to be globalized. This interface is used to simplify the globalization calls in
the middle end. We can check any globalization call that is only called by a
single thread in the team and replace it with a static shared memory buffer.
Reviewed By: jdoerfert
Differential Revision: https://reviews.llvm.org/D97818
Summary:
Memory globalization is required to maintain OpenMP standard semantics for data sharing between
worker and master threads. The GPU cannot share data between its threads so must allocate global or
shared memory to store the data in. Currently this is implemented fully in the frontend using the
`__kmpc_data_sharing_push_stack` and __kmpc_data_sharing_pop_stack` functions to emulate standard
CPU stack sharing. The front-end scans the target region for variables that escape the region and
must be shared between the threads. Each variable then has a field created for it in a global record
type.
This patch replaces this functinality with a single allocation command, effectively mimicing an
alloca instruction for the variables that must be shared between the threads. This will be much
slower than the current solution, but makes it much easier to optimize as we can analyze each
variable independently and determine if it is not captured. In the future, we can replace these
calls with an `alloca` and small allocations can be pushed to shared memory.
Reviewed By: tianshilei1992
Differential Revision: https://reviews.llvm.org/D97680
The initial use for AAExecutionDomain was to determine if a single
thread executes a block. While this is sometimes informative most
of the time, and for other reasons, we actually want to know if it
is the "initial thread". Thus, the thread that started execution on
the current device. The deduction needs to be adjusted in a follow
up as the methods we use right not are looking for the OpenMP thread
id which is resets whenever a thread enters a parallel region. What
we basically want is to look for `llvm.nvvm.read.ptx.sreg.ntid.x` and
equivalent functions.
<string> is currently the highest impact header in a clang+llvm build:
https://commondatastorage.googleapis.com/chromium-browser-clang/llvm-include-analysis.html
One of the most common places this is being included is the APInt.h header, which needs it for an old toString() implementation that returns std::string - an inefficient method compared to the SmallString versions that it actually wraps.
This patch replaces these APInt/APSInt methods with a pair of llvm::toString() helpers inside StringExtras.h, adjusts users accordingly and removes the <string> from APInt.h - I was hoping that more of these users could be converted to use the SmallString methods, but it appears that most end up creating a std::string anyhow. I avoided trying to use the raw_ostream << operators as well as I didn't want to lose having the integer radix explicit in the code.
Differential Revision: https://reviews.llvm.org/D103888
Summary:
Currently, only `OptimizationRemarks` can be emitted using a Function.
Add constructors to allow this for `OptimizationRemarksAnalysis` and
`OptimizationRemarkMissed` as well.
Reviewed By: jdoerfert thegameg
Differential Revision: https://reviews.llvm.org/D102784
Summary:
The OpenMP runtime functions don't always provide unique thread ID's to
determine if a basic block is truly single-threaded. Change the implementation
to only check NVPTX intrinsics for now.
Reviewed By: jdoerfert
Differential Revision: https://reviews.llvm.org/D102700
Summary:
This patch prevents the Attributor instances made in the CGSCC pass from
deleting functions. This prevents the attributor from changing the call
graph while OpenMPOpt is working with it.
Reviewed By: jdoerfert
Differential Revision: https://reviews.llvm.org/D102363
Summary:
Add the AAExecutionDomainInfo attributor instance to OpenMPOpt.
This will infer information relating to domain information that an
instruction might be expecting in. Right now this only includes a very
crude check for instructions that will be executed by the master thread
by comparing a thread-id function with a constant zero.
Reviewed By: jdoerfert
Differential Revision: https://reviews.llvm.org/D101578
This revision simplifies Clang codegen for parallel regions in OpenMP GPU target offloading and corresponding changes in libomptarget: SPMD/non-SPMD parallel calls are unified under a single `kmpc_parallel_51` runtime entry point for parallel regions (which will be commonized between target, host-side parallel regions), data sharing is internalized to the runtime. Tests have been auto-generated using `update_cc_test_checks.py`. Also, the revision contains changes to OpenMPOpt for remark creation on target offloading regions.
Reviewed By: jdoerfert, Meinersbur
Differential Revision: https://reviews.llvm.org/D95976
Summary:
This patch registers OpenMPOpt as a Module pass in addition to a CGSCC
pass. This is so certain optimzations that are sensitive to intact
call-sites can happen before inlining. The old `openmpopt` pass name is
changed to `openmp-opt-cgscc` and `openmp-opt` calls the Module pass.
The current module pass only runs a single check but will be expanded in
the future.
Reviewed By: jdoerfert
Differential Revision: https://reviews.llvm.org/D99202
Initial support for using the OpenMPIRBuilder by clang to generate loops using the OpenMPIRBuilder. This initial support is intentionally limited to:
* Only the worksharing-loop directive.
* Recognizes only the nowait clause.
* No loop nests with more than one loop.
* Untested with templates, exceptions.
* Semantic checking left to the existing infrastructure.
This patch introduces a new AST node, OMPCanonicalLoop, which becomes parent of any loop that has to adheres to the restrictions as specified by the OpenMP standard. These restrictions allow OMPCanonicalLoop to provide the following additional information that depends on base language semantics:
* The distance function: How many loop iterations there will be before entering the loop nest.
* The loop variable function: Conversion from a logical iteration number to the loop variable.
These allow the OpenMPIRBuilder to act solely using logical iteration numbers without needing to be concerned with iterator semantics between calling the distance function and determining what the value of the loop variable ought to be. Any OpenMP logical should be done by the OpenMPIRBuilder such that it can be reused MLIR OpenMP dialect and thus by flang.
The distance and loop variable function are implemented using lambdas (or more exactly: CapturedStmt because lambda implementation is more interviewed with the parser). It is up to the OpenMPIRBuilder how they are called which depends on what is done with the loop. By default, these are emitted as outlined functions but we might think about emitting them inline as the OpenMPRuntime does.
For compatibility with the current OpenMP implementation, even though not necessary for the OpenMPIRBuilder, OMPCanonicalLoop can still be nested within OMPLoopDirectives' CapturedStmt. Although OMPCanonicalLoop's are not currently generated when the OpenMPIRBuilder is not enabled, these can just be skipped when not using the OpenMPIRBuilder in case we don't want to make the AST dependent on the EnableOMPBuilder setting.
Loop nests with more than one loop require support by the OpenMPIRBuilder (D93268). A simple implementation of non-rectangular loop nests would add another lambda function that returns whether a loop iteration of the rectangular overapproximation is also within its non-rectangular subset.
Reviewed By: jdenny
Differential Revision: https://reviews.llvm.org/D94973
The existing implementation of parallel region merging applies only to
consecutive parallel regions that have speculatable sequential
instructions in-between. This patch lifts this limitation to expand
merging with any sequential instructions in-between, except calls to
unmergable OpenMP runtime functions. In-between sequential instructions
in the merged region are sequentialized in a "master" region and any
output values are broadcasted to the following parallel regions and the
sequential region continuation of the merged region.
Reviewed By: jdoerfert
Differential Revision: https://reviews.llvm.org/D90909
Previously when trying to support CoroSplit's function splitting, we
added in a hack that simply added the new function's node into the
original function's SCC (https://reviews.llvm.org/D87798). This is
incorrect since it might be in its own SCC.
Now, more similar to the previous design, we have callers explicitly
notify the LazyCallGraph that a function has been split out from another
one.
In order to properly support CoroSplit, there are two ways functions can
be split out.
One is the normal expected "outlining" of one function into a new one.
The new function may only contain references to other functions that the
original did. The original function must reference the new function. The
new function may reference the original function, which can result in
the new function being in the same SCC as the original function. The
weird case is when the original function indirectly references the new
function, but the new function directly calls the original function,
resulting in the new SCC being a parent of the original function's SCC.
This form of function splitting works with CoroSplit's Switch ABI.
The second way of splitting is more specific to CoroSplit. CoroSplit's
Retcon and Async ABIs split the original function into multiple
functions that all reference each other and are referenced by the
original function. In order to keep the LazyCallGraph in a valid state,
all new functions must be processed together, else some nodes won't be
populated. To keep things simple, this only supports the case where all
new edges are ref edges, and every new function references every other
new function. There can be a reference back from any new function to the
original function, putting all functions in the same RefSCC.
This also adds asserts that all nodes in a (Ref)SCC can reach all other
nodes to prevent future incorrect hacks.
The original hacks in https://reviews.llvm.org/D87798 are no longer
necessary since all new functions should have been registered before
calling updateCGAndAnalysisManagerForPass.
This fixes all coroutine tests when opt's -enable-new-pm is true by
default. This also fixes PR48190, which was likely due to the previous
hack breaking SCC invariants.
Reviewed By: rnk
Differential Revision: https://reviews.llvm.org/D93828
If a GPU function is externally reachable we give up trying to find the
(unique) kernel it is called from. This can hinder optimizations. Emit a
remark and explain mitigation strategies.
Reviewed By: tianshilei1992
Differential Revision: https://reviews.llvm.org/D93439
The OpenMP 5.1 assumptions `no_openmp` and `no_openmp_routines` allow us
to ignore calls that would otherwise prevent ICV tracking.
Once we track more ICVs we might need to distinguish the ones that could
be impacted even with `no_openmp_routines`.
Reviewed By: sstefan1
Differential Revision: https://reviews.llvm.org/D92050
OpenMPIRBuilder::createParallel outlines the body region of the parallel
construct into a new function that accepts any value previously defined outside
the region as a function argument. This function is called back by OpenMP
runtime function __kmpc_fork_call, which expects trailing arguments to be
pointers. If the region uses a value that is not of a pointer type, e.g. a
struct, the produced code would be invalid. In such cases, make createParallel
emit IR that stores the value on stack and pass the pointer to the outlined
function instead. The outlined function then loads the value back and uses as
normal.
Reviewed By: jdoerfert, llitchev
Differential Revision: https://reviews.llvm.org/D92189
Summary:
Add support for passing source locations to libomptarget runtime functions using the ident_t struct present in the rest of the libomp API. This will allow the runtime system to give much more insightful error messages and debugging values.
Reviewers: jdoerfert grokos
Differential Revision: https://reviews.llvm.org/D87946
Summary:
This patch adds support for passing in the original delcaration name in the source file to the libomptarget runtime. This will allow the runtime to provide more intelligent debugging messages. This patch takes the original expression parsed from the OpenMP map / update clause and provides a textual representation if it was explicitly mapped, otherwise it takes the name of the variable declaration as a fallback. The information in passed to the runtime in a global array of strings that matches the existing ident_t source location strings using ";name;filename;column;row;;"
Reviewers: jdoerfert
Differential Revision: https://reviews.llvm.org/D89802
For consistency with the IRBuilder, OpenMPIRBuilder has method names starting with 'Create'. However, the LLVM coding style has methods names starting with lower case letters, as all other OpenMPIRBuilder already methods do. The clang-tidy configuration used by Phabricator also warns about the naming violation, adding noise to the reviews.
This patch renames all `OpenMPIRBuilder::CreateXYZ` methods to `OpenMPIRBuilder::createXYZ`, and updates all in-tree callers.
I tested check-llvm, check-clang, check-mlir and check-flang to ensure that I did not miss a caller.
Reviewed By: mehdi_amini, fghanim
Differential Revision: https://reviews.llvm.org/D91109
Summary:
This patch adds support for passing in the original delcaration name in the
source file to the libomptarget runtime. This will allow the runtime to provide
more intelligent debugging messages. This patch takes the original expression
parsed from the OpenMP map / update clause and provides a textual
representation if it was explicitly mapped, otherwise it takes the name of the
variable declaration as a fallback. The information in passed to the runtime in
a global array of strings that matches the existing ident_t source location
strings using ";name;filename;column;row;;". See
clang/test/OpenMP/target_map_names.cpp for an example of the generated output
for a given map clause.
Reviewers: jdoervert
Differential Revision: https://reviews.llvm.org/D89802
There are cases that generated OpenMP code consists of multiple,
consecutive OpenMP parallel regions, either due to high-level
programming models, such as RAJA, Kokkos, lowering to OpenMP code, or
simply because the programmer parallelized code this way. This
optimization merges consecutive parallel OpenMP regions to: (1) reduce
the runtime overhead of re-activating a team of threads; (2) enlarge the
scope for other OpenMP optimizations, e.g., runtime call deduplication
and synchronization elimination.
This implementation defensively merges parallel regions, only when they
are within the same BB and any in-between instructions are safe to
execute in parallel.
Reviewed By: jdoerfert
Differential Revision: https://reviews.llvm.org/D83635
Summary:
This patch add support for printing analysis messages relating to data
globalization on the GPU. This occurs when data is shared between the
threads in a GPU context and must be pushed to global or shared memory.
Reviewers: jdoerfert
Subscribers: guansong hiraditya llvm-commits ormris sstefan1 yaxunl
Tags: #OpenMP #LLVM
Differential Revision: https://reviews.llvm.org/D88243
Refactored __tgt_target_data_begin_mapper_<issue|wait> to receive the handle as an input/output argument.
This given the compiler warning of returning the handle as copy.
Differential Revision: https://reviews.llvm.org/D88029
When checking call sites, give special handling to indirect call, as the
callee may be unknown and can lead to nullptr dereference later. Assume
conservatively that the ICV always changes in such case.
Reviewed By: sstefan1
Differential Revision: https://reviews.llvm.org/D87104
getValuesInOffloadArrays goes through the offload arrays in __tgt_target_data_begin_mapper getting the values stored in them before the call is issued.
call void @__tgt_target_data_begin_mapper(arg0, arg1,
i8** %offload_baseptrs, i8** %offload_ptrs, i64* %offload_sizes,
...)
Diferential Revision: https://reviews.llvm.org/D86300
The problem with module slice has been addressed in D86319
Introduce two new AAs. AAICVTrackerFunctionReturned which checks if a
function can have a unique ICV value after it is finished, and
AAICVCallSiteReturned which checks AAICVTrackerFunctionReturned for a
call site. This enables us to check the value of a call and if it
changes the ICV. This also changes the approach in
`getReplacementValues()` to a worklist-based approach so we can explore
all relevant BBs.
Differential Revision: https://reviews.llvm.org/D85544
Summary:
The module slice describes which functions we can analyze and transform
while working on an SCC as part of the Attributor-CGSCC pass. So far we
simply restricted it to the SCC.
Reviewers: jdoerfert
Differential Revision: https://reviews.llvm.org/D86319
This commits breaks certain OpenMP codes (on power) because it expanded
the Attributor scope without telling the Attributor about the SCC
extend. See: https://reviews.llvm.org/D85544#2227611
This reverts commit b0b32e649011d9a60165b9b53eb2764b7da9c8ca.
canBeMovedDownwards checks if the "wait" counterpart of __tgt_target_data_begin_mapper can be moved downwards, returning a pointer to the instruction that might require/modify the data transferred, and returning null it the movement is not possible or not worth it. The function splitTargetDataBeginRTC receives that returned instruction and instead of moving the "wait" it creates it at that point.
Differential Revision: https://reviews.llvm.org/D86155
Introduce two new AAs. AAICVTrackerFunctionReturned which checks if a
function can have a unique ICV value after it is finished, and
AAICVCallSiteReturned which checks AAICVTrackerFunctionReturned for a
call site. This enables us to check the value of a call and if it
changes the ICV. This also changes the approach in
`getReplacementValues()` to a worklist-based approach so we can explore
all relevant BBs.
Differential Revision: https://reviews.llvm.org/D85544
WIP that tries to hide the latency of runtime calls that involve host to
device memory transfers by splitting them into their "issue" and "wait"
versions. The "issue" is moved upwards as much as possible. The "wait" is
moved downards as much as possible. The "issue" issues the memory transfer
asynchronously, returning a handle. The "wait" waits in the returned
handle for the memory transfer to finish. We still lack of the movement.
Summary:
This seems obvious in hindsight, but the result is surprising.
I've measured compile-time of `-openmpopt` pass standalone
on RawSpeed unity build, and while there is some OpenMP stuff,
most is not OpenMP. But nonetheless the pass does a lot of costly
preparations before ever trying to look for OpenMP stuff in SCC.
Numbers (n=25): 0.094624s -> 0.005976s, an -93.68% improvement, or ~16x
Reviewers: jdoerfert
Reviewed By: jdoerfert
Subscribers: yaxunl, hiraditya, guansong, llvm-commits, sstefan1
Tags: #llvm
Differential Revision: https://reviews.llvm.org/D84689
Since D83271 we can optimize the GPU state machine to avoid spurious
call edges that increase the register usage of kernels. With this patch
we inform the user why and if this optimization is happening and when it
is not.
Reviewed By: ye-luo
Differential Revision: https://reviews.llvm.org/D83707
Summary: The `getIdAddr()` function returns the address of the ID of the abstract attribute
Reviewers: jdoerfert, sstefan1, uenoku, homerdin, baziotis
Reviewed By: jdoerfert
Subscribers: okura, hiraditya, uenoku, kuter, llvm-commits
Tags: #llvm
Differential Revision: https://reviews.llvm.org/D83172
