Skip creating a module if no function is going to be imported.
Also includes a change so that if the first partition is empty (which
can happen),
we import global with non-local linkage into the first non-empty
partition, instead
of P0 all the time.
I thought we'd need to change users of the SplitModule callback so they
can deal with less modules
than the number requested, but no. We already return only 1 module in
some cases and
it seems to be handled just fine.
Fixes SWDEV-523146
Fix a reference to getValue() being optional in InlineSizeEstimatorAnalysis, a
file that is not included in the default build. A "warning: enumerated and
non-enumerated type in conditional expression" warning is fixed in AMDGPU too.
InstructionCost is already an optional value, containing an Invalid
state that can be checked with isValid(). There is little point in
returning another optional from getValue(). Most uses do not make use of
it being a std::optional, dereferencing the value directly (either
isValid has been checked previously or the Cost is assumed to be valid).
The one case that does in AMDGPU used value_or which has been replaced
by a isValid() check.
Introduce members() iterator-helper to shorten the members_{begin,end}
idiom. A previous attempt of this patch was #130319, which had to be
reverted due to unit-test failures when attempting to call members() on
the end iterator. In this patch, members() accepts either an ECValue or
an ElemTy, which is more intuitive and doesn't suffer from the same
issue.
Currently iterators over EquivalenceClasses will iterate over std::set,
which guarantees the order specified by the comperator. Unfortunately in
many cases, EquivalenceClasses are used with pointers, so iterating over
std::set of pointers will not be deterministic across runs.
There are multiple places that explicitly try to sort the equivalence
classes before using them to try to get a deterministic order
(LowerTypeTests, SplitModule), but there are others that do not at the
moment and this can result at least in non-determinstic value naming in
Float2Int.
This patch updates EquivalenceClasses to keep track of all members via a
extra SmallVector and removes code from LowerTypeTests and SplitModule
to sort the classes before processing.
Overall it looks like compile-time slightly decreases in most cases, but
close to noise:
https://llvm-compile-time-tracker.com/compare.php?from=7d441d9892295a6eb8aaf481e1715f039f6f224f&to=b0c2ac67a88d3ef86987e2f82115ea0170675a17&stat=instructions
PR: https://github.com/llvm/llvm-project/pull/134075
DenseSet, SmallPtrSet, SmallSet, SetVector, and StringSet recently
gained C++23-style insert_range. This patch replaces:
Dest.insert(Src.begin(), Src.end());
with:
Dest.insert_range(Src);
This patch does not touch custom begin like succ_begin for now.
This PR abstracts the `LockFileManager` API into new `AdvisoryLock`
interface. This is so that we can create an alternative implementation
for Clang implicitly-built modules that is optimized for single-process
environment.
Clients of `LockFileManager` call `unsafeRemoveLockFile()` whenever
`tryLock()` fails. However looking at the code, there are no scenarios
where this actually does something useful. This PR removes such calls.
This patch removes some internal state out of `LockFileManager` by
moving the locking code from the constructor into new member function
`tryLock()` which returns the errors right away. This simplifies and
modernizes the interface.
Module split assumes that a kernel function must have an external
linkage; however, that isn't the case. For example, a static kernel
function will have a weak_odr linkage
Change-Id: I1e5dee0de1fd866b365f4090a574e1b2961f8dca
A static analysis tool warned that a division was always being performed
in integer division, so was either 0.0 or 1.0.
This doesn't seem intentional, so has been fixed to return a true ratio
using floating-point division. This in turn showed a bug where a
comparison against this ratio was incorrect.
A static analysis tool found that ModuleCost could be zero, so would
perform divide by zero when being printed. Perhaps this is unreachable
in practice, but the fix is straightforward enough and unlikely to be a
performance concern.
(reland with fixed sed command for macos)
Handle the `!callees` metadata to further reduce the amount of indirect
call cases that end up conservatively assuming that any indirectly
callable function is a potential target.
See #106528 to review the first commit.
Handle the `!callees` metadata to further reduce the amount of indirect
call cases that end up conservatively assuming that any indirectly
callable function is a potential target.
- Don't treat inline ASM as indirect calls
- Remove call to alias testing, which was broken (only working by pure
luck right now) and isn't needed anyway. GlobalOpt should take care of
them for us.
Relands #104763 with
- Fixes for EXPENSIVE_CHECKS test failure (due to sorting operator
failing if the input is shuffled first)
- Fix for broken proposal selection
- c3cb27370af40e491446164840766478d3258429 included
Original commit description below
---
Major rewrite of the AMDGPUSplitModule pass in order to better support
it long-term.
Highlights:
- Removal of the "SML" logging system in favor of just using CL options
and LLVM_DEBUG, like any other pass in LLVM.
- The SML system started from good intentions, but it was too flawed and
messy to be of any real use. It was also a real pain to use and made the
code more annoying to maintain.
- Graph-based module representation with DOTGraph printing support
- The graph represents the module accurately, with bidirectional, typed
edges between nodes (a node usually represents one function).
- Nodes are assigned IDs starting from 0, which allows us to represent a
set of nodes as a BitVector. This makes comparing 2 sets of nodes to
find common dependencies a trivial task. Merging two clusters of nodes
together is also really trivial.
- No more defaulting to "P0" for external calls
- Roots that can reach non-copyable dependencies (such as external
calls) are now grouped together in a single "cluster" that can go into
any partition.
- No more defaulting to "P0" for indirect calls
- New representation for module splitting proposals that can be graded
and compared.
- Graph-search algorithm that can explore multiple branches/assignments
for a cluster of functions, up to a maximum depth.
- With the default max depth of 8, we can create up to 256 propositions
to try and find the best one.
- We can still fall back to a greedy approach upon reaching max depth.
That greedy approach uses almost identical heuristics to the previous
version of the pass.
All of this gives us a lot of room to experiment with new heuristics or
even entirely different splitting strategies if we need to. For
instance, the graph representation has room for abstract nodes, e.g. if
we need to represent some global variables or external constraints. We
could also introduce more edge types to model other type of relations
between nodes, etc.
I also designed the graph representation & the splitting strategies to
be as fast as possible, and it seems to have paid off. Some quick tests
showed that we spend pretty much all of our time in the CloneModule
function, with the actual splitting logic being >1% of the runtime.
Major rewrite of the AMDGPUSplitModule pass in order to better support
it long-term.
Highlights:
- Removal of the "SML" logging system in favor of just using CL options
and LLVM_DEBUG, like any other pass in LLVM.
- The SML system started from good intentions, but it was too flawed and
messy to be of any real use. It was also a real pain to use and made the
code more annoying to maintain.
- Graph-based module representation with DOTGraph printing support
- The graph represents the module accurately, with bidirectional, typed
edges between nodes (a node usually represents one function).
- Nodes are assigned IDs starting from 0, which allows us to represent a
set of nodes as a BitVector. This makes comparing 2 sets of nodes to
find common dependencies a trivial task. Merging two clusters of nodes
together is also really trivial.
- No more defaulting to "P0" for external calls
- Roots that can reach non-copyable dependencies (such as external
calls) are now grouped together in a single "cluster" that can go into
any partition.
- No more defaulting to "P0" for indirect calls
- New representation for module splitting proposals that can be graded
and compared.
- Graph-search algorithm that can explore multiple branches/assignments
for a cluster of functions, up to a maximum depth.
- With the default max depth of 8, we can create up to 256 propositions
to try and find the best one.
- We can still fall back to a greedy approach upon reaching max depth.
That greedy approach uses almost identical heuristics to the previous
version of the pass.
All of this gives us a lot of room to experiment with new heuristics or
even entirely different splitting strategies if we need to. For
instance, the graph representation has room for abstract nodes, e.g. if
we need to represent some global variables or external constraints. We
could also introduce more edge types to model other type of relations
between nodes, etc.
I also designed the graph representation & the splitting strategies to
be as fast as possible, and it seems to have paid off. Some quick tests
showed that we spend pretty much all of our time in the CloneModule
function, with the actual splitting logic being >1% of the runtime.
I initially assumed only kernels could be roots, but that is wrong. A
function with no callers also needs to be a root to ensure it is
correctly handled.
They're very rare because we usually internalize everything, and
internal functions with no callers would be deleted.
When they are present, we need to also consider their dependencies and
act accordingly. Previously, we could put a function "by default" in P0,
but it could call another function with internal linkage defined in
another module which was of course incorrect.
Fixes SWDEV-467695
It allows it to access TTI correctly, and opens the door to accessing
more analysis in the future.
I went back and forth between this, and also making the default
SplitModule a Pass too to make it uniform, but I decided against it
because it's just needless complications. Neither llvm-split or
LTOBackend have a PM ready to use so we need to create one anyway. Let's
keep all the mess hidden in the AMDGPU version for now to keep this
change more self-contained.
This is just something I noticed while going over this pass logic one
more time and didn't cause issues (yet). If we find an indirect call, we
stop looking assuming we added all functions to the list, but if not all
functions in the module were indirectly callable, some may still be
missing.
Just to be safe, keep looking until we did everything we could to find
dependencies, so we don't accidentally miss one.
(with fix for ubsan)
This enables the --lto-partitions option to work more consistently.
This module splitting logic is fully aware of AMDGPU modules and their
specificities and takes advantage of
them to split modules in a way that avoids compilation issue (such as
resource usage being incorrectly represented).
This also includes a logging system that's more elaborate than just
LLVM_DEBUG which allows
printing logs to uniquely named files, and optionally with all value
names hidden so they can be safely shared without leaking informatiton
about the source. Logs can also be enabled through an environment
variable, which avoids the sometimes complicated process of passing a
-mllvm option all the way from clang driver to the offload linker that
handles full LTO codegen.
This enables the --lto-partitions option to work more consistently.
This module splitting logic is fully aware of AMDGPU modules and their
specificities and takes advantage of
them to split modules in a way that avoids compilation issue (such as
resource usage being incorrectly represented).
This also includes a logging system that's more elaborate than just
LLVM_DEBUG which allows
printing logs to uniquely named files, and optionally with all value
names hidden so they can be safely shared without leaking informatiton
about the source. Logs can also be enabled through an environment
variable, which avoids the sometimes complicated process of passing a
-mllvm option all the way from clang driver to the offload linker that
handles full LTO codegen.
