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
Global initializers may be ConstantArrays. They need to be checked
explicitly, because different-yet-still-equivalent type names may be
used for each, and/or a GEP instruction may appear in one.
The only wrinkle is that we can't process the "blockaddress" arguments
of the callbr until the blocks have been equated. So we force them to be
"unified" before checking.
This was left out when the callbr instruction was added.
Differential Revision: https://reviews.llvm.org/D104606
When the load type is changed to ptr, we need the load pointer type
to also be ptr, because it's not allowed to create a pointer to an
opaque pointer. This is achieved by adjusting the getPointerTo() API
to return an opaque pointer for an opaque pointer base type.
Differential Revision: https://reviews.llvm.org/D104718
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
These used to consistently be zeroed pre-gfx9, but gfx9 made the
situation complicated since now some still do and some don't. This
also manages to pick up a few cases that the pattern fails to optimize
away.
We handle some cases with instruction patterns, but some get
through. In particular this improves the integer cases.
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:
The changes to globalization introduced in D97680 created two new functions to
push / pop shareably memory on the GPU, __kmpc_alloc_shared and
__kmpc_free_shared. This patch adds these new runtime functions to the
library info so they can be used by the HeapToStack attributor interface. This
optimization replaces malloc / free pairs with stack memory if legal.
Reviewed By: tianshilei1992
Differential Revision: https://reviews.llvm.org/D102087
0 latency instructions now get processed and retired properly within the in-order pipeline. Had to fix a bug within TimelineView.cpp as well that would show up when a 0 latency instruction was the first instruction in the source.
Differential Revision: https://reviews.llvm.org/D104675
We can do this optimization in the majority of cases, but we currently
don't have a way to do it. We do not track/model which instructions
have which behavior, the control bit to change the high bit behavior,
or making use of preserved bits at all. This is a bit fuzzy since we
don't know precisely how the source instruction will be lowered, but
that only really matters in one case (for fma_mixlo).
We do need to fixup some of these cases after selection, but the
pattern helps eliminate many of these zexts.
Summary:
generate eh_info when vector registers are saved according to the traceback table.
struct eh_info_t {
unsigned version; /* EH info version 0 */
#if defined(64BIT)
char _pad[4]; /* padding */
#endif
unsigned long lsda; /* Pointer to Language Specific Data Area */
unsigned long personality; /* Pointer to the personality routine */
};
the value of lsda and personality is zero when the number of vector registers saved is large zero and there is not personality of the function
Reviewers: Jason Liu
Differential Revision: https://reviews.llvm.org/D103651
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
Since Xcode 13, ld64 requires linking libSystem for all the executable.
Fix the tests that needs to run ld64 by linking libSystem from sysroot.
rdar://77332728
Reviewed By: fhahn
Differential Revision: https://reviews.llvm.org/D104332
Some APIs work with const variables while others don't. This can cause
conflicts when calling one from the other.
This is NFC.
Differential Revision: https://reviews.llvm.org/D104719
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
This adds support for addrspace casts involving opaque pointers to
InstCombine, as well as the isEliminableCastPair() helper
(otherwise the assertion failure would just move there).
Add PointerType::hasSameElementTypeAs() to hide the element type
details.
Differential Revision: https://reviews.llvm.org/D104668
This is a recommit that fixes unwanted STP generation by checking that
the base register has not been modified or used elsewhere.
Our initial motivating case was memcpy's with alignments > 16. The
loads/stores, to which small memcpy's expand, are kept together in
several places so that we get a sequence like this for a 64 bit copy:
LD w0
LD w1
ST w0
ST w1
The load/store optimiser can generate a LDP/STP w0, w1 from this because
the registers read/written are consecutive. In our case however, the
sequence is optimised during ISel, resulting in:
LD w0
ST w0
LD w0
ST w0
This instruction reordering allows reuse of registers. Since the registers
are no longer consecutive (i.e. they are the same), it inhibits LDP/STP
creation. The approach here is to perform renaming:
LD w0
ST w0
LD w1
ST w1
to enable the folding of the stores into a STP. We do not yet generate
the LDP due to a limitation in the renaming implementation, but plan to
look at that in a follow-up so that we fully support this case. While
this was initially motivated by certain memcpy's, this is a general
approach and thus is beneficial for other cases too, as can be seen
in some test changes.
Differential Revision: https://reviews.llvm.org/D103597
There was a bug from cost calculation for partially invariant unswitch.
The costs of non-duplicated blocks are substracted from the total LoopCost, so
anything that is duplicated should not be counted.
Differential Revision: https://reviews.llvm.org/D103816
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
Don't do this while stipping pointer casts, instead fetch it at
the end. This improves compatibility with opaque pointers for the
case where the base object is not opaque.
Currently we drop wrapping flags for expressions like (A + C1)<flags> - C2.
But we can retain flags under certain conditions:
* Adding a smaller constant is NUW if the original AddExpr was NUW.
* Adding a constant with the same sign and small magnitude is NSW, if the
original AddExpr was NSW.
This can improve results after using `SimplifyICmpOperands`, which may
subtract one in order to use stricter predicates, as is the case for
`isKnownPredicate`.
Reviewed By: efriedma
Differential Revision: https://reviews.llvm.org/D104319
These serve as a convenient combination of consume_front/back and
startswith_lower/endswith_lower, consistent with other existing
case insensitive methods named <operation>_lower.
Differential Revision: https://reviews.llvm.org/D104218
Handle to gep p, 0-v case separately, and not as part of the loop
that ensures all indices are constant integers. Those two things
are not really related.
This patch adds the `afn`, `contract`, and `reassoc` fast-math flags.
It also fixes up `fneg`'s order in the alphabetized list.
Reviewed By: MaskRay, craig.topper
Differential Revision: https://reviews.llvm.org/D104541
This patch aims to add the scalable property to LLT. The rest of the
patch-series changes the interfaces to take/return ElementCount and
TypeSize, which both have the ability to represent the scalable property.
The changes are mostly mechanical and aim to be non-functional changes
for fixed-width vectors.
For scalable vectors some unit tests have been added, but no effort has
been put into making any of the GlobalISel algorithms work with scalable
vectors yet. That will be left as future work.
The work is split into a series of 5 patches to make reviews easier.
Reviewed By: arsenm
Differential Revision: https://reviews.llvm.org/D104450
