Add an interface that allows grouping together all covolution and
pooling ops within Linalg named ops. The interface currently
- the indexing map used for input/image access is valid
- the filter and output are accessed using projected permutations
- that all loops are charecterizable as one iterating over
- batch dimension,
- output image dimensions,
- filter convolved dimensions,
- output channel dimensions,
- input channel dimensions,
- depth multiplier (for depthwise convolutions)
Differential Revision: https://reviews.llvm.org/D109793
This partially reverts commits 1fc2a47f0b and 9816e726e7.
See D109727. Replacing config.guess in favor of {gcc,clang} -dumpmachine
can avoid the riscv64-{redhat,suse}-linux GCC detection.
Acked-by: Luís Marques <luismarques@lowrisc.org>
All supported compilers have supported `=delete` as an extension
in C++03 mode for many years at this point.
Differential Revision: https://reviews.llvm.org/D109942
When adding an image to a target for crashlog purposes, avoid specifying
the architecture of the image.
This has the effect of making SBTarget::AddModule infer the ArchSpec for
the image based on the SBTarget's architecture, which LLDB puts serious
effort into calculating correctly (in TargetList::CreateTargetInternal).
The status quo is that LLDB randomly guesses the ArchSpec for a module
if its architecture is specified, via:
```
SBTarget::AddModule -> Platform::GetAugmentedArchSpec -> Platform::IsCompatibleArchitecture ->
GetSupportedArchitectureAtIndex -> {ARM,x86}GetSupportedArchitectureAtIndex
```
... which means that the same crashlog can fail to load on an Apple
Silicon Mac (due to the random guess of arm64e-apple-macosx for the
module's ArchSpec not being compatible with the SBTarget's (correct)
ArchSpec), while loading just fine on an Intel Mac.
I'm not sure how to add a test for this (it doesn't look like there's
test coverage of this path in-tree). It seems like it would be pretty
complicated to regression test: the host LLDB would need to be built for
arm64e, we'd need a hand-crafted arm64e iOS crashlog, and we'd need a
binary with an iOS deployment target. I'm open to other / simpler
options.
rdar://82679400
Differential Revision: https://reviews.llvm.org/D110013
Similar to D69607 but for archive member extraction unrelated to GC. This patch adds --why-extract=.
Prior art:
GNU ld -M prints
```
Archive member included to satisfy reference by file (symbol)
a.a(a.o) main.o (a)
b.a(b.o) (b())
```
-M is mainly for input section/symbol assignment <-> output section mapping
(often huge output) and the information may appear ad-hoc.
Apple ld64
```
__Z1bv forced load of b.a(b.o)
_a forced load of a.a(a.o)
```
It doesn't say the reference file.
Arm's proprietary linker
```
Selecting member vsnprintf.o(c_wfu.l) to define vsnprintf.
...
Loading member vsnprintf.o from c_wfu.l.
definition: vsnprintf
reference : _printf_a
```
---
--why-extract= gives the user the full data (which is much shorter than GNU ld
-Map). It is easy to track a chain of references to one archive member with a
one-liner, e.g.
```
% ld.lld main.o a_b.a b_c.a c.a -o /dev/null --why-extract=- | tee stdout
reference extracted symbol
main.o a_b.a(a_b.o) a
a_b.a(a_b.o) b_c.a(b_c.o) b()
b_c.a(b_c.o) c.a(c.o) c()
% ruby -ane 'BEGIN{p={}}; p[$F[1]]=[$F[0],$F[2]] if $.>1; END{x="c.a(c.o)"; while y=p[x]; puts "#{y[0]} extracts #{x} to resolve #{y[1]}"; x=y[0] end}' stdout
b_c.a(b_c.o) extracts c.a(c.o) to resolve c()
a_b.a(a_b.o) extracts b_c.a(b_c.o) to resolve b()
main.o extracts a_b.a(a_b.o) to resolve a
```
Archive member extraction happens before --gc-sections, so this may not be a live path
under --gc-sections, but I think it is a good approximation in practice.
* Specifying a file avoids output interleaving with --verbose.
* Required `=` prevents accidental overwrite of an input if the user forgets `=`. (Most of compiler drivers' long options accept `=` but not ` `)
Differential Revision: https://reviews.llvm.org/D109572
Some buildbots fail with:
> C:\a\llvm-clang-x86_64-expensive-checks-win\llvm-project\llvm\lib\IR\Verifier.cpp(4352): error C2678: binary '==': no operator found which takes a left-hand operand of type 'const llvm::MDOperand' (or there is no acceptable conversion)
Possibly the explicit MDOperand to Metadata* conversion will help?
This patch fixes the warning
InstructionTables.cpp:27:56: error: loop variable 'Resource' of type
'const std::pair<const uint64_t, ResourceUsage> &' (aka 'const
pair<const unsigned long, llvm::mca::ResourceUsage> &') binds to a
temporary constructed from type 'const std::pair<unsigned long,
llvm::mca::ResourceUsage> &' [-Werror,-Wrange-loop-construct]
Note that Resource is declared as:
SmallVector<std::pair<uint64_t, ResourceUsage>, 4> Resources;
without "const" for uint64_t.
For strided accesses the loop vectorizer seems to prefer creating a
vector induction variable with a start value of the form
<i32 0, i32 1, i32 2, ...>. This value will be incremented each
loop iteration by a splat constant equal to the length of the vector.
Within the loop, arithmetic using splat values will be done on this
vector induction variable to produce indices for a vector GEP.
This pass attempts to dig through the arithmetic back to the phi
to create a new scalar induction variable and a stride. We push
all of the arithmetic out of the loop by folding it into the start,
step, and stride values. Then we create a scalar GEP to use as the
base pointer for a strided load or store using the computed stride.
Loop strength reduce will run after this pass and can do some
cleanups to the scalar GEP and induction variable.
Reviewed By: frasercrmck
Differential Revision: https://reviews.llvm.org/D107790
This moves the logic for adding symbols based on UUID, file and frame
into little helper functions. This is in preparation for D110011.
Differential revision: https://reviews.llvm.org/D110010
Adds additional tests following comments from D109844.
Also removes unusued in.ptr arguments and places in the call tests that
used loads instead of a getval call.
This pass transforms SCF.ForOp operations to SCF.WhileOp. The For loop condition is placed in the 'before' region of the while operation, and indctuion variable incrementation + the loop body in the 'after' region. The loop carried values of the while op are the induction variable (IV) of the for-loop + any iter_args specified for the for-loop.
Any 'yield' ops in the for-loop are rewritten to additionally yield the (incremented) induction variable.
This transformation is useful for passes where we want to consider structured control flow solely on the basis of a loop body and the computation of a loop condition. As an example, when doing high-level synthesis in CIRCT, the incrementation of an IV in a for-loop is "just another part" of a circuit datapath, and what we really care about is the distinction between our datapath and our control logic (the condition variable).
Differential Revision: https://reviews.llvm.org/D108454
Reworked reordering algorithm. Originally, the compiler just tried to
detect the most common order in the reordarable nodes (loads, stores,
extractelements,extractvalues) and then fully rebuilding the graph in
the best order. This was not effecient, since it required an extra
memory and time for building/rebuilding tree, double the use of the
scheduling budget, which could lead to missing vectorization due to
exausted scheduling resources.
Patch provide 2-way approach for graph reodering problem. At first, all
reordering is done in-place, it doe not required tree
deleting/rebuilding, it just rotates the scalars/orders/reuses masks in
the graph node.
The first step (top-to bottom) rotates the whole graph, similarly to the previous
implementation. Compiler counts the number of the most used orders of
the graph nodes with the same vectorization factor and then rotates the
subgraph with the given vectorization factor to the most used order, if
it is not empty. Then repeats the same procedure for the subgraphs with
the smaller vectorization factor. We can do this because we still need
to reshuffle smaller subgraph when buildiong operands for the graph
nodes with lasrger vectorization factor, we can rotate just subgraph,
not the whole graph.
The second step (bottom-to-top) scans through the leaves and tries to
detect the users of the leaves which can be reordered. If the leaves can
be reorder in the best fashion, they are reordered and their user too.
It allows to remove double shuffles to the same ordering of the operands in
many cases and just reorder the user operations instead. Plus, it moves
the final shuffles closer to the top of the graph and in many cases
allows to remove extra shuffle because the same procedure is repeated
again and we can again merge some reordering masks and reorder user nodes
instead of the operands.
Also, patch improves cost model for gathering of loads, which improves
x264 benchmark in some cases.
Gives about +2% on AVX512 + LTO (more expected for AVX/AVX2) for {625,525}x264,
+3% for 508.namd, improves most of other benchmarks.
The compile and link time are almost the same, though in some cases it
should be better (we're not doing an extra instruction scheduling
anymore) + we may vectorize more code for the large basic blocks again
because of saving scheduling budget.
Differential Revision: https://reviews.llvm.org/D105020
Some intrinsic functions weren't findable because the table
wasn't strictly in order of names.
And complete a missing generalization of the extension DCONJG
to accept any kind of complex argument, like DREAL and DIMAG
were.
Differential Revision: https://reviews.llvm.org/D110002
D109607 results in a regression in llvm-test-suite.
The reason is we didn't check the size of SourceTy, so that we will
return wrong SSE type when SourceTy is overlapped.
Reviewed By: Meinersbur
Differential Revision: https://reviews.llvm.org/D110037
Atomics in C++ for OpenCL 2021 are now handled the same way as in
OpenCL C 3.0. This is a header-only change.
Differential Revision: https://reviews.llvm.org/D109424
Add a new version of fusion on tensors that supports the following scenarios:
- support input and output operand fusion
- fuse a producer result passed in via tile loop iteration arguments (update the tile loop iteration arguments)
- supports only linalg operations on tensors
- supports only scf::for
- cannot add an output to the tile loop nest
The LinalgTileAndFuseOnTensors pass tiles the root operation and fuses its producers.
Reviewed By: nicolasvasilache, mravishankar
Differential Revision: https://reviews.llvm.org/D109766
The docs of alpha.cplusplus.SmartPtr was incorrectly placed under
alpha.deadcode. Moved it to under alpha.cplusplus
Differential Revision: https://reviews.llvm.org/D110032
In ValueTracking.cpp we use a function called
computeKnownBitsFromOperator to determine the known bits of a value.
For the vscale intrinsic if the function contains the vscale_range
attribute we can use the maximum and minimum values of vscale to
determine some known zero and one bits. This should help to improve
code quality by allowing certain optimisations to take place.
Tests added here:
Transforms/InstCombine/icmp-vscale.ll
Differential Revision: https://reviews.llvm.org/D109883
Following D109516, this patch re-uses the new helper function for ELF relocation traversal in the RISCV backend.
Reviewed By: StephenFan
Differential Revision: https://reviews.llvm.org/D109522
Following D109516, this patch re-uses the new helper function for ELF relocation traversal in the x86-64 backend.
Reviewed By: StephenFan
Differential Revision: https://reviews.llvm.org/D109520
Previous changes like D101202 and D104261 have eliminated the special
status that break and continue once had, since now we're making
decisions purely based on the structure of the CFG without regard for
the underlying source code constructs.
This means we don't gain anything from defering handling for these
blocks. Dropping it moves some diagnostics, though arguably into a
better place. We're working around a "quirk" in the CFG that perhaps
wasn't visible before: while loops have an empty "transition block"
where continue statements and the regular loop exit meet, before
continuing to the loop entry. To get a source location for that, we
slightly extend our handling for empty blocks. The source location for
the transition ends up to be the loop entry then, but formally this
isn't a back edge. We pretend it is anyway. (This is safe: we can always
treat edges as back edges, it just means we allow less and don't modify
the lock set. The other way around it wouldn't be safe.)
Reviewed By: aaron.ballman
Differential Revision: https://reviews.llvm.org/D106715
Set value_regs and invalidate_regs in RegisterInfo pushed onto m_regs
to nullptr, to ensure that the temporaries passed there are not
accidentally used.
Differential Revision: https://reviews.llvm.org/D109879
The vectorizer can sometimes make reverse shuffles from indices that
count down. In MVE, we don't have a 128bit rev instruction, but we can
select this to a VREV64 with some lane movs to swap the two halfs.
Ideally this would use VMOVD's, but only gets as far as VMOVS's at the
moment.
Differential Revision: https://reviews.llvm.org/D69510
Previously the script emitted output using plain CHECK directives. This
can result in a test passing even if there are some instructions between
CHECK directives that should have been removed. It also makes debugging
tests that have the output in a different order more difficult since
FileCheck can match with a later line and then complain about the "wrong"
directive not being found.
This will cause quite large diffs when updating existing tests, but I'm not sure we need an opt-in flag here.
Depends on D109765 (pre-commit tests)
Reviewed By: MaskRay
Differential Revision: https://reviews.llvm.org/D109767
Since https://reviews.llvm.org/D87118, the StaticAnalyzer directory is
added unconditionally. In theory this should not cause the static analyzer
sources to be built unless they are referenced by another target. However,
the clang-cpp target (defined in clang/tools/clang-shlib) uses the
CLANG_STATIC_LIBS global property to determine which libraries need to
be included. To solve this issue, this patch avoids adding libraries to
that property if EXCLUDE_FROM_ALL is set.
In case something like this comes up again: `cmake --graphviz=targets.dot`
is quite useful to see why a target is included as part of `ninja all`.
Reviewed By: thakis
Differential Revision: https://reviews.llvm.org/D109611
Make use of runtime extension for the second reference counter used in
structured data region. This extension is implemented in D106510 and D106509.
Differential Revision: https://reviews.llvm.org/D106517
Always send PID in the detach packet when multiprocess extensions are
enabled. This is required by qemu's GDB server, as plain 'D' packet
results in an error and the emulated system is not resumed.
Differential Revision: https://reviews.llvm.org/D110033