This patch adds parallel processing of chunks. When reducing very large
inputs, e.g. functions with 500k basic blocks, processing chunks in
parallel can significantly speed up the reduction.
To allow modifying clones of the original module in parallel, each clone
needs their own LLVMContext object. To achieve this, each job parses the
input module with their own LLVMContext. In case a job successfully
reduced the input, it serializes the result module as bitcode into a
result array.
To ensure parallel reduction produces the same results as serial
reduction, only the first successfully reduced result is used, and
results of other successful jobs are dropped. Processing resumes after
the chunk that was successfully reduced.
The number of threads to use can be configured using the -j option.
It defaults to 1, which means serial processing.
Reviewed By: Meinersbur
Differential Revision: https://reviews.llvm.org/D113857
This patch moves the logic to clone and check a new chunk into a new
function, to allow re-use in a follow-up patch that implements parallel
reductions.
Reviewed By: dblaikie
Differential Revision: https://reviews.llvm.org/D113856
Textual LLVM IR files are much bigger and take longer to write to disk.
To avoid the extra cost incurred by serializing to text, this patch adds
an option to save temporary files as bitcode instead.
Reviewed By: aeubanks
Differential Revision: https://reviews.llvm.org/D113858
Having a separate counting method runs the risk of a mismatch between
the actual reduction method and the counting method.
Instead, create an Oracle that always returns true for shouldKeep(), run
the reduction, and count how many times shouldKeep() was called. The
module should not be modified if shouldKeep() always returns true.
Reviewed By: Meinersbur
Differential Revision: https://reviews.llvm.org/D113537
Metadata operands tend to require special conditions, especially on dbg
intrinsics. We also don't have a zero value for metadata.
Replacing callee operands is a little weird, since calling undef/null
doesn't make sense. It also causes tons of invalid reductions when
reducing calls to intrinsics since only arguments to intrinsics can be
of the metadata type.
Reviewed By: Meinersbur
Differential Revision: https://reviews.llvm.org/D113532
Add a new "operands-skip" pass whose goal is to remove instructions in the middle of dependency chains. For instance:
```
%baseptr = alloca i32
%arrayidx = getelementptr i32, i32* %baseptr, i32 %idxprom
store i32 42, i32* %arrayidx
```
might be reducible to
```
%baseptr = alloca i32
%arrayidx = getelementptr ... ; now dead, together with the computation of %idxprom
store i32 42, i32* %baseptr
```
Other passes would either replace `%baseptr` with undef (operands, instructions) or move it to become a function argument (operands-to-args), both of which might fail the interestingness check.
In principle the implementation allows operand replacement with any value or instruction in the function that passes the filter constraints (same type, dominance, "more reduced"), but is limited in this patch to values that are directly or indirectly used to compute the current operand value, motivated by the example above. Additionally, function arguments are added to the candidate set which helps reducing the number of relevant arguments mitigating a concern of too many arguments mentioned in https://reviews.llvm.org/D110274#3025013.
Possible future extensions:
* Instead of requiring the same type, bitcast/trunc/zext could be automatically inserted for some more flexibility.
* If undef is added to the candidate set, "operands-skip"is able to produce any reduction that "operands" can do. Additional candidates might be zero and one, where the "reductive power" classification can prefer one over the other. If undefined behaviour should not be introduced, undef can be removed from the candidate set.
Recommit after resolving conflict with D112651 and reusing
shouldReduceOperand from D113532.
Reviewed By: aeubanks
Differential Revision: https://reviews.llvm.org/D111818
Add a new "operands-skip" pass whose goal is to remove instructions in the middle of dependency chains. For instance:
```
%baseptr = alloca i32
%arrayidx = getelementptr i32, i32* %baseptr, i32 %idxprom
store i32 42, i32* %arrayidx
```
might be reducible to
```
%baseptr = alloca i32
%arrayidx = getelementptr ... ; now dead, together with the computation of %idxprom
store i32 42, i32* %baseptr
```
Other passes would either replace `%baseptr` with undef (operands, instructions) or move it to become a function argument (operands-to-args), both of which might fail the interestingness check.
In principle the implementation allows operand replacement with any value or instruction in the function that passes the filter constraints (same type, dominance, "more reduced"), but is limited in this patch to values that are directly or indirectly used to compute the current operand value, motivated by the example above. Additionally, function arguments are added to the candidate set which helps reducing the number of relevant arguments mitigating a concern of too many arguments mentioned in https://reviews.llvm.org/D110274#3025013.
Possible future extensions:
* Instead of requiring the same type, bitcast/trunc/zext could be automatically inserted for some more flexibility.
* If undef is added to the candidate set, "operands-skip"is able to produce any reduction that "operands" can do. Additional candidates might be zero and one, where the "reductive power" classification can prefer one over the other. If undefined behaviour should not be introduced, undef can be removed from the candidate set.
Reviewed By: aeubanks
Differential Revision: https://reviews.llvm.org/D111818
When reducing functions with very large basic blocks (~ almost 1 million
BBs), the majority of time is spent maintaining the order in the std::set
for the basic blocks to keep.
In those cases, DenseSet<> is much more efficient. Use it instead.
Previously, if the basic-blocks delta pass tried to remove a basic block
that was the last basic block in a function that did not have external
or weak linkage, the resulting IR would become invalid. Since removing
the last basic block in a function is effectively identical to removing
the function body itself, we check explicitly for this case and if we
detect it, we run the same logic as in ReduceFunctionBodies.cpp
Reviewed By: aeubanks
Differential Revision: https://reviews.llvm.org/D113486
Sometimes if llvm-reduce is interrupted in the middle of a delta pass on
a large file, it can take quite some time for the tool to start actually
doing new work if it is restarted again on the partially-reduced file. A
lot of time ends up being spent testing large chunks when these large
chunks are very unlikely to actually pass the interestingness test. In
cases like this, the tool will complete faster if the starting
granularity is reduced to a finer amount. Thus, we introduce a command
line flag that automatically divides the chunks into smaller subsets a
fixed, user-specified number of times prior to beginning the core loop.
Reviewed By: aeubanks
Differential Revision: https://reviews.llvm.org/D112651
(Second try. Need to link against CodeGen and MC libs.)
The llvm-reduce tool has been extended to operate on MIR (import, clone and
export). Current limitation is that only a single machine function is
supported. A single reducer pass that operates on machine instructions (while
on SSA-form) has been added. Additional MIR specific reducer passes can be
added later as needed.
Differential Revision: https://reviews.llvm.org/D110527
The llvm-reduce tool has been extended to operate on MIR (import, clone and
export). Current limitation is that only a single machine function is
supported. A single reducer pass that operates on machine instructions (while
on SSA-form) has been added. Additional MIR specific reducer passes can be
added later as needed.
Differential Revision: https://reviews.llvm.org/D110527
The extractBasicBlocksFromModule, extractInstrFromModule, and other
similar functions previously performed very poorly when the number of
such elements in the program to reduce was very high. Previously, we
were creating the set which caches elements to keep by looping through
all elements in the module and adding them to the set. However, since
std::set is an ordered set, this introduces a massive amount of
rebalancing if the order of elements in the program and the order of
their pointers in memory are not the same.
The solution is straightforward: first put all the elements to be kept
in a vector, then use the constructor for std::set which takes a pair of
iterators over a collection. This constructor is optimized to avoid
doing unnecessary work when initializing large sets.
Also in this change, we pass BBsToKeep set to functions
replaceBranchTerminator and removeUninterestingBBsFromSwitch as a const
reference rather than passing it by value. This ought to prevent the
need to copy the collection each time these functions are called, which
is expensive if the collection is large.
Reviewed By: aeubanks
Differential Revision: https://reviews.llvm.org/D112757
This was checked while counting but not actually when doing the reduction, resulting in crashes.
Reviewed By: Meinersbur
Differential Revision: https://reviews.llvm.org/D112766
Having non-undef constants in a final llvm-reduce output is nicer than
having undefs.
This splits the existing reduce-operands pass into three, one which does
the same as the current pass of reducing to undef, and two more to
reduce to the constant 1 and the constant 0. Do not reduce to undef if
the operand is a ConstantData, and do not reduce 0s to 1s.
Reducing GEP operands very frequently causes invalid IR (since types may
not match up if we index differently into a struct), so don't touch GEPs.
Reviewed By: Meinersbur
Differential Revision: https://reviews.llvm.org/D111765
Instead of setting operands to undef as the "operands" pass does,
convert the operands to a function argument. This avoids having to
introduce undef values into the IR which have some unpredictability
during optimizations.
For instance,
define void @func() {
entry:
%val = add i32 32, 21
store i32 %val, i32* null
ret void
}
is reduced to
define void @func(i32 %val) {
entry:
%val1 = add i32 32, 21
store i32 %val, i32* null
ret void
}
(note that the instruction %val is renamed to %val1 when printing
the IR to avoid ambiguity; ideally %val1 would be removed by dce or the
instruction reduction pass)
Any call to @func is replaced with a call to the function with the
new signature and filled with undef. This is not ideal for IPA passes,
but those out-of-scope for now.
Reviewed By: aeubanks
Differential Revision: https://reviews.llvm.org/D111503
Use Module& wherever possible.
Since every reduction immediately turns Chunks into an Oracle, directly pass Oracle instead.
Reviewed By: hans
Differential Revision: https://reviews.llvm.org/D111122
We expose the fact that we rely on unsigned wrapping to iterate through
all indexes. This can be confusing. Rather, keeping it as an
implementation detail through an iterator is less confusing and is less
code.
Reviewed By: rnk
Differential Revision: https://reviews.llvm.org/D110885
When replacing function calls, skip call instructions where the old
function is not the called function, but e.g. the old function is passed
as an argument.
This fixes a crash due to trying to construct invalid IR for the test
case.
Reviewed By: aeubanks
Differential Revision: https://reviews.llvm.org/D109759
The ReduceMetadata pass before this patch removed metadata on a per-MDNode (or NamedMDNode) basis. Either all references to an MDNode are kept, or all of them are removed. However, MDNodes are uniqued, meaning that references to MDNodes with the same data become references to the same MDNodes. As a consequence, e.g. tbaa references to the same type will all have the same MDNode reference and hence make it impossible to reduce only keeping metadata on those memory access for which they are interesting.
Moreover, MDNodes can also be referenced by some intrinsics or other MDNodes. These references were not considered for removal leading to the possibility that MDNodes are not actually removed even if selected to be removed by the oracle.
This patch changes ReduceMetadata to reduces based on removable metadata references instead. MDNodes without references implicitly dropped anyway. References by intrinsic calls should be removed by ReduceOperands or ReduceInstructions. References in other MDNodes cannot be removed as it would violate the immutability of MDNodes.
Additionally, ReduceMetadata pass before this patch used `setMetadata(I, NULL)` to remove references, where `I` is the index in the array returned by `getAllMetadata`. However, `setMetadata` expects a MDKind (such as `MD_tbaa`) as first argument. `getAllMetadata` does not return those in consecutive order (otherwise it would not need to be a `std::pair` with `first` representing the MDKind).
Reviewed By: aeubanks, swamulism
Differential Revision: https://reviews.llvm.org/D110534
This removes the data layout, target triple, source filename, and module
identifier when possible.
Reviewed By: swamulism
Differential Revision: https://reviews.llvm.org/D108568
This patch allows iterating typed enum via the ADT/Sequence utility.
It also changes the original design to better separate concerns:
- `StrongInt` only deals with safe `intmax_t` operations,
- `SafeIntIterator` presents the iterator and reverse iterator
interface but only deals with safe `StrongInt` internally.
- `iota_range` only deals with `SafeIntIterator` internally.
This design ensures that operations are always valid. In particular,
"Out of bounds" assertions fire when:
- the `value_type` is not representable as an `intmax_t`
- iterator operations make internal computation underflow/overflow
- the internal representation cannot be converted back to `value_type`
Differential Revision: https://reviews.llvm.org/D106279
The argument reduction pass shouldn't remove arguments of
intrinsics, because the resulting module is ill-formed, and so
inherently uninteresting.
Reviewed By: aeubanks
Differential Revision: https://reviews.llvm.org/D103129
The parseInputFile function returns an empty unique_ptr to signal an
error, like when the input file doesn't exist, or is malformed. In this
case, the tool should exit immediately rather than segfault by
dereferencing the unique_ptr later.
Reviewed By: aeubanks
Differential Revision: https://reviews.llvm.org/D102891
This introduces a flag that aborts if we ever reduce to IR that fails
the verifier.
Reviewed By: swamulism, arichardson
Differential Revision: https://reviews.llvm.org/D101279
Much like with ReduceFunctionBodies delta pass,
we need to remove comdat and set linkage to external,
else verifier will complain, and our deltas are invalid.
The limitation of the current pass that it skips initializer-less GV's
seems arbitrary, in all the reduced cases i (personally) looked at,
the globals weren't needed, yet they were kept.
So let's do two things:
1. allow reducing initializer-less globals
2. before reducing globals, reduce their initializers, much like we do function bodies
ee6e25e439 changed
the delta pass to skip intrinsics, which means we may end up being
left with declarations of intrinsics, that aren't otherwise referenced
in the module. This is obviously unwanted, do drop them.
No longer rely on an external tool to build the llvm component layout.
Instead, leverage the existing `add_llvm_componentlibrary` cmake function and
introduce `add_llvm_component_group` to accurately describe component behavior.
These function store extra properties in the created targets. These properties
are processed once all components are defined to resolve library dependencies
and produce the header expected by llvm-config.
Differential Revision: https://reviews.llvm.org/D90848
This patch adds a reduction of 'special' globals that lead to further
reductions (e.g. alias or regular globals reduction) being less efficient
because there are special constraints on values referenced in those
special globals. For example, values in @llvm.used and
@llvm.compiler.used need to be named, so replacing all uses of an
alias/global with undef or a different unnamed constant results in
invalid IR.
More details:
https://llvm.org/docs/LangRef.html#intrinsic-global-variables
Reviewed By: lebedev.ri
Differential Revision: https://reviews.llvm.org/D90302
This patch adds a new reduction pass that tries to remove aliases.
It runs early, as most of those likely can be removed up-front in
practice.
This substantially improves llvm-reduce for IR generated by the swift
compiler, which can generate a lot of aliases which lead to lots of
invalid reductions.
Reviewed By: lebedev.ri
Differential Revision: https://reviews.llvm.org/D90260
Florian Hahn