Performance on GPU targets can be highly variable, sometimes inlining
everything hurts performance and sometimes it greatly improves it. Add
an option to toggle this behaviour to better investigate it.
Reviewed By: jdoerfert
Differential Revision: https://reviews.llvm.org/D109014
In LLVM IR, `AlignmentBitfieldElementT` is 5-bit wide
But that means that the maximal alignment exponent is `(1<<5)-2`,
which is `30`, not `29`. And indeed, alignment of `1073741824`
roundtrips IR serialization-deserialization.
While this doesn't seem all that important, this doubles
the maximal supported alignment from 512MiB to 1GiB,
and there's actually one noticeable use-case for that;
On X86, the huge pages can have sizes of 2MiB and 1GiB (!).
So while this doesn't add support for truly huge alignments,
which i think we can easily-ish do if wanted, i think this adds
zero-cost support for a not-trivially-dismissable case.
I don't believe we need any upgrade infrastructure,
and since we don't explicitly record the IR version,
we don't need to bump one either.
As @craig.topper speculates in D108661#2963519,
this might be an artificial limit imposed by the original implementation
of the `getAlignment()` functions.
Differential Revision: https://reviews.llvm.org/D108661
Use uint64_t for lanemask on all GPU architectures at the interface
with clang. Updates tests. The deviceRTL is always linked as IR so the zext
and trunc introduced for wave32 architectures will fold after inlining.
Simplification partly motivated by amdgpu gfx10 which will be wave32 and
is awkward to express in the current arch-dependant typedef interface.
Reviewed By: jdoerfert
Differential Revision: https://reviews.llvm.org/D108317
Currently, AAKernelInfo will fail on an assertion if we attempt to run
it on a kernel without the init / deinit runtime calls. However, this
occurs for global constructors on the device. This will cause OpenMPOpt
to crash whenever global constructors are present. This patch removes
this assertion and just gives up instead.
Reviewed By: jdoerfert
Differential Revision: https://reviews.llvm.org/D108258
To avoid simplification with wrong constants we need to make sure we
know that we won't perform specific optimizations based on the users
request. The non-SPMDzation and non-CustomStateMachine flags did only
prevent the final transformation but allowed to value simplification
to go ahead.
Differential Revision: https://reviews.llvm.org/D107862
This patch adds the `AlwaysInline` attribute to the `__kmpc_parallel_51`
device runtime call. This improves inlining heuristics which encourages
the indirect function pointer arguemnt to also be inlined. This greatly
improves performance for a few applications whose outlined regions were
not inlined otherwise.
Reviewed By: jdoerfert
Differential Revision: https://reviews.llvm.org/D107839
This patch expands SPMDization (converting generic execution mode to SPMD for target regions) by guarding code regions that should be executed only by the main thread. Specifically, it generates guarded regions, which only the main thread executes, and the synchronization with worker threads using simple barriers. For correctness, the patch aborts SPMDization for target regions if the same code executes in a parallel region, thus must be not be guarded. This check is implemented using the ParallelLevels AA.
Reviewed By: jhuber6
Differential Revision: https://reviews.llvm.org/D106892
This work provides four flags to disable four different sets of OpenMP optimizations. These flags take effect in llvm/lib/Transforms/IPO/OpenMPOpt.cpp and include the following:
- openmp-opt-disable-deglobalization: Defaults to false, adding this flag sets the variable DisableOpenMPOptDeglobalization to true. This prevents AA registration for HeapToStack and HeapToShared.
- openmp-opt-disable-spmdization: Defaults to false, adding this flag sets the variable DisableOpenMPOptSPMDization to true. This indicates a pessimistic fixpoint in changeToSPMDMode.
- openmp-opt-disable-folding: Defaults to false, adding this flag sets the variable DisableOpenMPOptFolding to true. This indicates a pessimistic fixpoint in the attributor init for AAFoldRuntimeCall.
- openmp-opt-disable-state-machine-rewrite: Defaults to false, adding this flag sets the variable DisableOpenMPOptStateMachineRewrite to true. This first prevents changes to the state machine in rewriteDeviceCodeStateMachine by returning before changes are made, and if a custom state machine is built in buildCustomStateMachine, stops by returning a pessimistic fixpoint.
Reviewed By: jhuber6
Differential Revision: https://reviews.llvm.org/D106802
The current implementation of function internalization creats a copy of each
function and replaces every use. This has the downside that the external
versions of the functions will call into the internalized versions of the
functions. This prevents them from being fully independent of eachother. This
patch replaces the current internalization scheme with a method that creates
all the copies of the functions intended to be internalized first and then
replaces the uses as long as their caller is not already internalized.
Reviewed By: jdoerfert
Differential Revision: https://reviews.llvm.org/D106931
The device runtime contains several calls to `__kmpc_get_hardware_num_threads_in_block`
and `__kmpc_get_hardware_num_blocks`. If the thread_limit and the num_teams are constant,
these calls can be folded to the constant value.
In this patch we use the already introduced `AAFoldRuntimeCall` and the `NumTeams` and
`NumThreads` kernel attributes (to be introduced in a different patch) to fold these functions.
The code checks all the kernels, and if their attributes match, the functions are folded.
In the future we will explore specializing for multiple values of NumThreads and NumTeams.
Depends on D106390
Reviewed By: jdoerfert, JonChesterfield
Differential Revision: https://reviews.llvm.org/D106033
D106185 allows us to determine if a store is needed easily. Using that
knowledge we can start to delete dead stores.
In AAIsDead we now track more state as an instruction can be dead (= the
old optimisitc state) or just "removable". A store instruction can be
removable while being very much alive, e.g., if it stores a constant
into an alloca or internal global. If we would pretend it was dead
instead of only removablewe we would ignore it when we determine what
values a load can see, so that is not what we want.
Differential Revision: https://reviews.llvm.org/D106188
This patch introduces `getPotentialCopiesOfStoredValue` which uses
AAPointerInfo to determine all "aliases" or "potential copies" of a
value that is stored into memory. This operation can fail but if it
succeeds it means we can visit all "uses" of a value even if it is
temporarily stored in memory.
There are two users for the function:
1) `Attributor::checkForAllUses` which will now ignore the value use
in a store if all "potential copies" can be identified and instead
be visited. This allows various AAs, including AAPointerInfo
itself, to look through memory.
2) `AANoCapture` which uses a custom use tracking through the
CaptureTracker interface and therefore needs to be thought
explicitly.
Differential Revision: https://reviews.llvm.org/D106185
Similar to D105787, this patch tries to fold `__kmpc_parallel_level` if possible.
Note that `__kmpc_parallel_level` doesn't take activeness into consideration,
based on current `deviceRTLs`, its return value can be such as 0, 1, 2, instead
of 0, 129, 130, etc. that also indicate activeness.
Reviewed By: jdoerfert
Differential Revision: https://reviews.llvm.org/D106154
While rewriteDeviceCodeStateMachine should probably be folded into
buildCustomStateMachine, we at least need the optimization to happen.
This was not reliably the case in the CGSCC pass but in the Module pass
it seems to work reliably.
This also ports a test to the new kernel encoding (target_init/deinit),
and makes sure we cannot run the kernel in SPMD mode.
Differential Revision: https://reviews.llvm.org/D106345
Deduplication in OpenMPOpt finds redundant OpenMP runtime calls and replaces them with a single call placed in the earliest safe location in the IR. When deduplication happens in a target region this patch makes sure replacement calls are put after target_init.
Reviewed By: jdoerfert
Differential Revision: https://reviews.llvm.org/D106556
SPMDization D102307 detects incompatible OpenMP runtime calls to abort converting a target region to SPMD mode. Calls to memory allocation/de-allocation routines kmpc_alloc_shared, kmpc_free_shared are incompatible unless they are removed by AAHeapToStack/AAHeapToShared analysis. This patch extends SPMDization detection to include AAHeapToStack/AAHeapToShared analysis results for enlarging the scope of possible SPMDized regions detected.
Reviewed By: jdoerfert
Differential Revision: https://reviews.llvm.org/D105634
This patch strips the NoInline attribute from known OpenMP runtime functions.
This is done so that we can denote certain runtime functions as NoInline to
ensure their call sites are intact so they can be checked by OpenMPOpt. We
don't wan't this noinline attribute to remain for any functions after OpenMPOpt
has been run however.
Reviewed By: jdoerfert
Differential Revision: https://reviews.llvm.org/D106482
This patch adds the ability to fold `__kmpc_is_generic_main_thread_id` if we
know for a fact that it is executed by the initial thread using
AAExecutionDomain. This combined with folding `__kmpc_is_spmd_exec_mode` will
allow us to fully fold `__kmpc_is_generic_main_thread`.
Depends on D106438 D106437
Reviewed By: jdoerfert
Differential Revision: https://reviews.llvm.org/D106439
Qualified kernels can be transformed from generic-mode to SPMD mode using an
optimization in OpenMPOpt. This patch introduces a new execution mode to
indicate kernels that have been transformed from generic-mode to SPMD-mode.
These kernels have SPMD-mode execution, but need generic-mode semantics for
scheduling the blocks and threads. Without this far too few blocks will be
scheduled for a generic region as SPMD mode expects the trip count to be
divided by the number of threads.
Reviewed By: ggeorgakoudis
Differential Revision: https://reviews.llvm.org/D106460
This patch changes `__kmpc_free_shared` to take an additional argument
corresponding to the associated allocation's size. This makes it easier to
implement the allocator in the runtime.
Reviewed By: jdoerfert
Differential Revision: https://reviews.llvm.org/D106496
SPMDization in D102307 does not change the RequiresFullRuntime argument of kmpc_target_init/deinit calls. However, the constraints of SPMDization detection for converting a target region to SPMD mode should guarantee that the region does not require full runtime support. Hence, this patch sets RequiresFullRuntime to false for improved execution performance.
Reviewed By: jdoerfert
Differential Revision: https://reviews.llvm.org/D105556
This patch introduces AAPointerInfo which tracks the uses of a pointer
and places them in "bins" based on their offset from the base and access
size.
As with other AAs, any pointer can be tracked but it is up to the user
to make sense of the results. The user in this patch is AAValueSimplify
and AAPotentialValues which both utilize AAPointerInfo to determine the
value of a load. For now, this is restricted to loads of allocas and
internal globals. Through the use of AAPointerInfo and the "bins" we can
track struct members separately. The users also know that storing only
zeros (at unknown indices) will result in loading only 0 (from unknown
indices). Other than that, the users are flow and context insensitive
(for now).
To deal with the "bins" more easily, AAPointerInfo provides a
forallInterfearingAccesses that applies a callback on all accesses
that might interfere with a given load or store.
Differential Revision: https://reviews.llvm.org/D104432
This patch rewrites and reworks a few of the existing remarks to make the mmore
concise and consistent prior to writing the documentation for them.
Reviewed By: jdoerfert
Differential Revision: https://reviews.llvm.org/D105898
In the device runtime there are many function calls to `__kmpc_is_spmd_exec_mode`
to query the execution mode of current kernels. In many cases, user programs
only contain target region executing in one mode. As a consequence, those runtime
function calls will only return one value. If we can get rid of these function
calls during compliation, it can potentially improve performance.
In this patch, we use `AAKernelInfo` to analyze kernel execution. Basically, for
each kernel (device) function `F`, we collect all kernel entries `K` that can
reach `F`. A new AA, `AAFoldRuntimeCall`, is created for each call site. In each
iteration, it will check all reaching kernel entries, and update the folded value
accordingly.
In the future we will support more function.
Reviewed By: jdoerfert
Differential Revision: https://reviews.llvm.org/D105787
In the device runtime there are many function calls to `__kmpc_is_spmd_exec_mode`
to query the execution mode of current kernels. In many cases, user programs
only contain target region executing in one mode. As a consequence, those runtime
function calls will only return one value. If we can get rid of these function
calls during compliation, it can potentially improve performance.
In this patch, we use `AAKernelInfo` to analyze kernel execution. Basically, for
each kernel (device) function `F`, we collect all kernel entries `K` that can
reach `F`. A new AA, `AAFoldRuntimeCall`, is created for each call site. In each
iteration, it will check all reaching kernel entries, and update the folded value
accordingly.
In the future we will support more function.
Reviewed By: jdoerfert
Differential Revision: https://reviews.llvm.org/D105787
In the spirit of TRegions [0], this patch analyzes a kernel and tracks
if it can be executed in SPMD-mode. If so, we flip the arguments of
the __kmpc_target_init and deinit call to enable the mode. We also
update the `<kernel>_exec_mode` flag to indicate to the runtime we
changed the mode to SPMD.
The code analysis is done interprocedurally by extending the
AAKernelInfo abstract attribute to track SPMD compatibility as well.
[0] https://link.springer.com/chapter/10.1007/978-3-030-28596-8_11
Differential Revision: https://reviews.llvm.org/D102307
In the spirit of TRegions [0], this patch creates a custom state
machine for a generic target region based on the potentially called
parallel regions.
The code analysis is done interprocedurally via an abstract attribute
(AAKernelInfo). All outermost parallel regions are collected and we
check if there might be unknown outermost parallel regions for which
we need an indirect call. Other AAKernelInfo extensions are expected.
[0] https://link.springer.com/chapter/10.1007/978-3-030-28596-8_11
Differential Revision: https://reviews.llvm.org/D101977
In the spirit of TRegions [0], this patch provides a simpler and uniform
interface for a kernel to set up the device runtime. The OMPIRBuilder is
used for reuse in Flang. A custom state machine will be generated in the
follow up patch.
The "surplus" threads of the "master warp" will not exit early anymore
so we need to use non-aligned barriers. The new runtime will not have an
extra warp but also require these non-aligned barriers.
[0] https://link.springer.com/chapter/10.1007/978-3-030-28596-8_11
This was in parts extracted from D59319.
Reviewed By: ABataev, JonChesterfield
Differential Revision: https://reviews.llvm.org/D101976
In order to simplify future extensions, e.g., the merge of
AAHeapToShared in to AAHeapToStack, we reorganize AAHeapToStack and the
state we keep for each malloc-like call. The result is also less
confusing as we only track malloc-like calls, not all calls. Further, we
only perform the updates necessary for a malloc-like to argue it can go
to the stack, e.g., we won't check all uses if we moved on to the
"must-be-freed" argument.
This patch also uses Attributor helps to simplify the allocated size,
alignment, and the potentially freed objects.
Overall, this is mostly a reorganization and only the use of the
optimistic helpers should change (=improve) the capabilities a bit.
Differential Revision: https://reviews.llvm.org/D104993
We should use AAValueSimplify for all value simplification, however
there was some leftover logic that predates AAValueSimplify in
AAReturnedValues. This remove the AAReturnedValues part and provides a
replacement by making AAValueSimplifyReturned strong enough to handle
all previously covered cases. Further, this improve
AAValueSimplifyCallSiteReturned to handle returned arguments.
AAReturnedValues is now much easier and the collected returned
values/instructions are now from the associated function only, making it
much more sane. We also do not have the brittle logic anymore that looks
for unresolved calls. Instead, we use AAValueSimplify to handle
recursion.
Useful code has been split into helper functions, e.g., an Attributor
interface to get a simplified value.
Differential Revision: https://reviews.llvm.org/D103860
Broke check-clang, see https://reviews.llvm.org/D102307#2869065
Ran `git revert -n ebbe149a6f08535ede848a531a601ae6591cfbc5..269416d41908bb670f67af689155d5ab8eea689a`
In the spirit of TRegions [0], this patch analyzes a kernel and tracks
if it can be executed in SPMD-mode. If so, we flip the arguments of
the __kmpc_target_init and deinit call to enable the mode. We also
update the `<kernel>_exec_mode` flag to indicate to the runtime we
changed the mode to SPMD.
The code analysis is done interprocedurally by extending the
AAKernelInfo abstract attribute to track SPMD compatibility as well.
[0] https://link.springer.com/chapter/10.1007/978-3-030-28596-8_11
Differential Revision: https://reviews.llvm.org/D102307
In the spirit of TRegions [0], this patch creates a custom state
machine for a generic target region based on the potentially called
parallel regions.
The code analysis is done interprocedurally via an abstract attribute
(AAKernelInfo). All outermost parallel regions are collected and we
check if there might be unknown outermost parallel regions for which
we need an indirect call. Other AAKernelInfo extensions are expected.
[0] https://link.springer.com/chapter/10.1007/978-3-030-28596-8_11
Differential Revision: https://reviews.llvm.org/D101977
In the spirit of TRegions [0], this patch provides a simpler and uniform
interface for a kernel to set up the device runtime. The OMPIRBuilder is
used for reuse in Flang. A custom state machine will be generated in the
follow up patch.
The "surplus" threads of the "master warp" will not exit early anymore
so we need to use non-aligned barriers. The new runtime will not have an
extra warp but also require these non-aligned barriers.
[0] https://link.springer.com/chapter/10.1007/978-3-030-28596-8_11
This was in parts extracted from D59319.
Reviewed By: ABataev, JonChesterfield
Differential Revision: https://reviews.llvm.org/D101976
In order to simplify future extensions, e.g., the merge of
AAHeapToShared in to AAHeapToStack, we reorganize AAHeapToStack and the
state we keep for each malloc-like call. The result is also less
confusing as we only track malloc-like calls, not all calls. Further, we
only perform the updates necessary for a malloc-like to argue it can go
to the stack, e.g., we won't check all uses if we moved on to the
"must-be-freed" argument.
This patch also uses Attributor helps to simplify the allocated size,
alignment, and the potentially freed objects.
Overall, this is mostly a reorganization and only the use of the
optimistic helpers should change (=improve) the capabilities a bit.
Differential Revision: https://reviews.llvm.org/D104993
We should use AAValueSimplify for all value simplification, however
there was some leftover logic that predates AAValueSimplify in
AAReturnedValues. This remove the AAReturnedValues part and provides a
replacement by making AAValueSimplifyReturned strong enough to handle
all previously covered cases. Further, this improve
AAValueSimplifyCallSiteReturned to handle returned arguments.
AAReturnedValues is now much easier and the collected returned
values/instructions are now from the associated function only, making it
much more sane. We also do not have the brittle logic anymore that looks
for unresolved calls. Instead, we use AAValueSimplify to handle
recursion.
Useful code has been split into helper functions, e.g., an Attributor
interface to get a simplified value.
Differential Revision: https://reviews.llvm.org/D103860
The remarks will trigger on some functions that are marked cold, such as the
`__muldc3` intrinsic functions. Change the remarks to avoid these functions.
Reviewed By: jdoerfert
Differential Revision: https://reviews.llvm.org/D105196
This patch adds additional remarks, suggesting the use of `noescape` for failed
globalization and indicating when internalization failed.
Reviewed By: jdoerfert
Differential Revision: https://reviews.llvm.org/D105150
Currently, LLParser will create a Function/GlobalVariable forward
reference based on the desired pointer type and then modify it when
it is declared. With opaque pointers, we generally do not know the
correct type to use until we see the declaration.
Solve this by creating the forward reference with a dummy type, and
then performing a RAUW with the correct Function/GlobalVariable when
it is declared. The approach is adopted from
b5b55963f6.
This results in a change to the use list order, which is why we see
test changes on some module passes that are not stable under use list
reordering.
Differential Revision: https://reviews.llvm.org/D104950
Currently OpenMPOpt will only check if a function is a kernel before deciding not to internalize it. Any uncalled function that gets internalized will be trivially dead in the module so this is unnnecessary.
Depends on D102423
Reviewed By: jdoerfert
Differential Revision: https://reviews.llvm.org/D104890
The metadata added in D102361 introduces a module flag that we can check
to determine if the module was compiled with `-fopenmp` enables. We can
now check for the precense of this instead of scanning the call graph
for OpenMP runtime functions.
Depends on D102361
Reviewed By: jdoerfert
Differential Revision: https://reviews.llvm.org/D102423
After landing the globalization optimizations, the precense of globalization on
the device that was not put in shared or stack memory is a failed optimization
with performance consequences so it should indicate a missed remark.
Reviewed By: jdoerfert
Differential Revision: https://reviews.llvm.org/D104735
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
