This feature is enabled by `-codegen-data-thinlto-two-rounds`, which
effectively runs the `-codegen-data-generate` and `-codegen-data-use` in
two rounds to enable global outlining with ThinLTO.
1. The first round: Run both optimization + codegen with a scratch
output.
Before running codegen, we serialize the optimized bitcode modules to a
temporary path.
2. From the scratch object files, we merge them into the codegen data.
3. The second round: Read the optimized bitcode modules and start the
codegen only this time.
Using the codegen data, the machine outliner effectively performs the
global outlining.
Depends on #90934, #110461 and #110463.
This is a patch for
https://discourse.llvm.org/t/rfc-enhanced-machine-outliner-part-2-thinlto-nolto/78753.
This is a prep for https://github.com/llvm/llvm-project/pull/90933.
- Change `ThinBackend` from a function to a type.
- Store the parallelism level in the type, which will be used when creating two-codegen round backends that inherit this value.
- `ThinBackendProc` is hoisted to `LTO.h` from `LTO.cpp` to provide its body for `ThinBackend`. However, `emitFiles()` is still implemented separately in `LTO.cpp`, distinct from its parent class.
We've noticed that for large builds executing thin-link can take on the
order of 10s of minutes. We are only using a single thread to write the
sharded indices and import files for each input bitcode file. While we
need to ensure the index file produced lists modules in a deterministic
order, that doesn't prevent us from executing the rest of the work in
parallel.
In this change we use a thread pool to execute as much of the backend's
work as possible in parallel. In local testing on a machine with 80
cores, this change makes a thin-link for ~100,000 input files run in ~2
minutes. Without this change it takes upwards of 10 minutes.
---------
Co-authored-by: Nuri Amari <nuriamari@fb.com>
This is a prep for https://github.com/llvm/llvm-project/pull/90933.
- Change `FileCache` from a function to a type.
- Store the cache directory in the type, which will be used when creating additional caches for two-codegen round runs that inherit this value.
Fix the use-after-free bug and re-apply
https://github.com/llvm/llvm-project/pull/106193
* Without the fix, the string referenced by `objSym.Name` could be
destroyed even if string saver keeps a copy of the referenced string.
This caused use-after-free.
* The fix ([latest
commit](9776ed44cf))
updates `objSym.Name` to reference (via `StringRef`) the string saver's
copy.
Test:
1. For `lld/test/ELF/lto/asmundef.ll`, its test failure is reproducible
with `-DLLVM_USE_SANITIZER=Address` and gone with the fix.
3. Run all tests by following
https://github.com/google/sanitizers/wiki/SanitizerBotReproduceBuild#try-local-changes.
* Without the fix, `ELF/lto/asmundef.ll` aborted the multi-stage test at
`@@@BUILD_STEP stage2/asan_ubsan check@@@`, defined
[here](https://github.com/llvm/llvm-zorg/blob/main/zorg/buildbot/builders/sanitizers/buildbot_fast.sh#L30)
* With the fix, the [multi-stage
test](https://github.com/llvm/llvm-zorg/blob/main/zorg/buildbot/builders/sanitizers/buildbot_fast.sh)
pass stage2 {asan, ubsan, masan}. This is also the test used by
https://lab.llvm.org/buildbot/#/builders/169
**Original commit message**
`StringMap<T>` creates a [copy of the
string](d4c519e7b2/llvm/include/llvm/ADT/StringMapEntry.h (L55-L58))
for entry insertions and intentionally keep copies [since the
implementation optimizes string memory
usage](d4c519e7b2/llvm/include/llvm/ADT/StringMap.h (L124)).
On the other hand, linker keeps copies of symbol names [1] in
`lld:🧝:parseFiles` [2] before invoking `compileBitcodeFiles` [3].
This change proposes to optimize away string copies inside
[LTO::GlobalResolutions](24e791b416/llvm/include/llvm/LTO/LTO.h (L409)),
which will make LTO indexing more memory efficient for ELF. There are
similar opportunities for other (COFF, wasm, MachO) formats.
The optimization takes place for lld (ELF) only. For the rest of use
cases (gold plugin, `llvm-lto2`, etc), LTO owns a string saver to keep
copies and use global resolution key for de-duplication.
Together with @kazutakahirata's work to make `ComputeCrossModuleImport`
more memory efficient, we see a ~20% peak memory usage reduction in a
binary where peak memory usage needs to go down. Thanks to the
optimization in
329ba523cc,
the max (as opposed to the sum) of `ComputeCrossModuleImport` or
`GlobalResolution` shows up in peak memory usage.
* Regarding correctness, the set of
[resolved](80c47ad3ae/llvm/lib/LTO/LTO.cpp (L739))
[per-module
symbols](80c47ad3ae/llvm/include/llvm/LTO/LTO.h (L188-L191))
is a subset of
[llvm::lto::InputFile::Symbols](80c47ad3ae/llvm/include/llvm/LTO/LTO.h (L120)).
And bitcode symbol parsing saves symbol name when iterating
`obj->symbols` in `BitcodeFile::parse` already. This change updates
`BitcodeFile::parseLazy` to keep copies of per-module undefined symbols.
* Presumably the undefined symbols in a LTO unit (copied in this patch
in linker unique saver) is a small set compared with the set of symbols
in global-resolution (copied before this patch), making this a
worthwhile trade-off. Benchmarking this change alone shows measurable
memory savings across various benchmarks.
[1] ELF
1cea5c2138/lld/ELF/InputFiles.cpp (L1748)
[2]
ef7b18a53c/lld/ELF/Driver.cpp (L2863)
[3]
ef7b18a53c/lld/ELF/Driver.cpp (L2995)
`StringMap<T>` creates a [copy of the
string](d4c519e7b2/llvm/include/llvm/ADT/StringMapEntry.h (L55-L58))
for entry insertions and intentionally keep copies [since the
implementation optimizes string memory
usage](d4c519e7b2/llvm/include/llvm/ADT/StringMap.h (L124)).
On the other hand, linker keeps copies of symbol names [1] in
`lld:🧝:parseFiles` [2] before invoking `compileBitcodeFiles` [3].
This change proposes to optimize away string copies inside
[LTO::GlobalResolutions](24e791b416/llvm/include/llvm/LTO/LTO.h (L409)),
which will make LTO indexing more memory efficient for ELF. There are
similar opportunities for other (COFF, wasm, MachO) formats.
The optimization takes place for lld (ELF) only. For the rest of use
cases (gold plugin, `llvm-lto2`, etc), LTO owns a string saver to keep
copies and use global resolution key for de-duplication.
Together with @kazutakahirata's work to make `ComputeCrossModuleImport`
more memory efficient, we see a ~20% peak memory usage reduction in a
binary where peak memory usage needs to go down. Thanks to the
optimization in
329ba523cc,
the max (as opposed to the sum) of `ComputeCrossModuleImport` or
`GlobalResolution` shows up in peak memory usage.
* Regarding correctness, the set of
[resolved](80c47ad3ae/llvm/lib/LTO/LTO.cpp (L739))
[per-module
symbols](80c47ad3ae/llvm/include/llvm/LTO/LTO.h (L188-L191))
is a subset of
[llvm::lto::InputFile::Symbols](80c47ad3ae/llvm/include/llvm/LTO/LTO.h (L120)).
And bitcode symbol parsing saves symbol name when iterating
`obj->symbols` in `BitcodeFile::parse` already. This change updates
`BitcodeFile::parseLazy` to keep copies of per-module undefined symbols.
* Presumably the undefined symbols in a LTO unit (copied in this patch
in linker unique saver) is a small set compared with the set of symbols
in global-resolution (copied before this patch), making this a
worthwhile trade-off. Benchmarking this change alone shows measurable
memory savings across various benchmarks.
[1] ELF
1cea5c2138/lld/ELF/InputFiles.cpp (L1748)
[2]
ef7b18a53c/lld/ELF/Driver.cpp (L2863)
[3]
ef7b18a53c/lld/ELF/Driver.cpp (L2995)
The primary motivation is to remove `EntryCount` from `FunctionSummary`.
This frees 8 bytes out of `sizeof(FunctionSummary)` (136 bytes as of
64498c5483).
While I'm at it, this PR clean up {SummaryBasedOptimizations,
SyntheticCountsPropagation} since they were not used and there are no
plans to further invest on them.
With this patch, bitcode writer writes a placeholder 0 at the byte
offset of `EntryCount` and bitcode reader can parse the function entry
count at the correct byte offset. Added a TODO to stop writing
`EntryCount` and bump bitcode version
Nothing went wrong in this case, we just successfully matched a module
by identifier. No need to print to std::error like we would for
something that should be user-visible.
Signed-off-by: Sarnie, Nick <nick.sarnie@intel.com>
This patch reduces the memory usage for import lists by employing
memory-efficient data structures.
With this patch, an import list for a given destination module is
basically DenseSet<uint32_t> with each element indexing into the
deduplication table containing tuples of:
{SourceModule, GUID, Definition/Declaration}
In one of our large applications, the peak memory usage goes down by
9.2% from 6.120GB to 5.555GB during the LTO indexing step.
This patch addresses several sources of space inefficiency associated
with std::unordered_map:
- std::unordered_map<GUID, ImportKind> takes up 16 bytes because of
padding even though ImportKind only carries one bit of information.
- std::unordered_map uses pointers to elements, both in the hash table
proper and for collision chains.
- We allocate an instance of std::unordered_map for each
{Destination Module, Source Module} pair for which we have at least
one import. Most import lists have less than 10 imports, so the
metadata like the size of std::unordered_map and the pointer to the
hash table costs a lot relative to the actual contents.
The background is as follows. I'm planning to reduce the memory
footprint of ThinLTO indexing by changing ImportMapTy, the data
structure used for an import list. Once this patch lands, I'm
planning to change the type slightly. The new type alias allows us to
update the type without touching many places.
This patch introduces type alias ModuleToSummariesForIndexTy.
I'm planning to change the type slightly to allow heterogeneous lookup
(that is, std::map<K, V, std::less<>>) in a subsequent patch. The
problem is that changing the type affects many places. Using a type
alias reduces the impact.
This patch turns type alias ImportMapTy into a proper class to provide
a more intuitive interface like:
ImportList.addDefinition(...)
as opposed to:
FunctionImporter::addDefinition(ImportList, ...)
Also, this patch requires all non-const accesses to go through
addDefinition, maybeAddDeclaration, and addGUID while providing const
accesses via:
const ImportMapTyImpl &getImportMap() const { return ImportMap; }
I realize ImportMapTy may not be the best name as a class (maybe OK as
a type alias). I am not renaming ImportMapTy in this patch at least
because there are 47 mentions of ImportMapTy under llvm/.
The two instances of std::set are used only for membership checking
purposes in computeLTOCacheKey. We do not need std::set's strengths
like iterators staying valid or the ability to traverse in a sorted
order. This patch changes them to DenseSet.
While I am at it, this patch replaces count with contains for slightly
increased readability.
Without this patch, computeLTOCacheKey computes SHA1, creates its
hexadecimal representation with toHex, which returns std::string, and
then copies it to an output parameter of type SmallString.
This patch removes the redirection and teaches computeLTOCacheKey to
directly return std::string computed by toHex. With the move
semantics, no buffer copy should be involved.
While I am at it, this patch adds a Twine to concatenate two strings.
This reverts commit 740161a9b98c9920dedf1852b5f1c94d0a683af5.
I moved the `ISD` dependencies into the CodeGen portion of the handling,
it's a little awkward but it's the easiest solution I can think of for
now.
Summary:
The LTO pass and LLD linker have logic in them that forces extraction
and prevent internalization of needed runtime calls. However, these
currently take all RTLibcalls into account, even if the target does not
support them. The target opts-out of a libcall if it sets its name to
nullptr. This patch pulls this logic out into a class in the header so
that LTO / lld can use it to determine if a symbol actually needs to be
kept.
This is important for targets like AMDGPU that want to be able to use
`lld` to perform the final link step, but does not want the overhead of
uncalled functions. (This adds like a second to the link time trivially)
https://github.com/llvm/llvm-project/pull/87600 was reverted in order to
revert
6262763341.
Now https://github.com/llvm/llvm-project/pull/95482 is fix forward for
6262763341.
This patch is a reland for
https://github.com/llvm/llvm-project/pull/87600
**Changes on top of original patch**
In `llvm/include/llvm/IR/ModuleSummaryIndex.h`, make the type of
`GVSummaryPtrSet` an `unordered_set` which is more memory efficient when
the number of elements is smaller than 128 [1]
**Original commit message**
For distributed ThinLTO, the LTO indexing step generates combined
summary for each module, and postlink pipeline reads the combined
summary which stores the information for link-time optimization.
This patch populates the 'import type' of a summary in bitcode, and
updates bitcode reader to parse the bit correctly.
[1]
393eff4e02/llvm/lib/Support/SmallPtrSet.cpp (L43)
https://github.com/llvm/llvm-project/pull/95482 is a reland of
https://github.com/llvm/llvm-project/pull/88024.
https://github.com/llvm/llvm-project/pull/95482 keeps indexing memory
usage reasonable by using unordered_map and doesn't make other changes
to originally reviewed code.
While discussing possible ways to minimize indexing memory usage, Teresa
asked whether I need `ExportSetTy` as a map or a set is sufficient. This
PR implements the idea. It uses a set rather than a map to track exposed
ValueInfos.
Currently, `ExportLists` has two use cases, and neither needs to track a
ValueInfo's import/export status. So using a set is sufficient and
correct.
1) In both in-process and distributed ThinLTO, it's used to decide if a
function or global variable is visible [1] from another module after importing
creates additional cross-module references.
* If a cross-module call edge is seen today, the callee must be visible
to another module without keeping track of its export status already.
For instance, this [2] is how callees of direct calls get exported.
2) For in-process ThinLTO [3], it's used to compute lto cache key.
* The cache key computation already hashes [4] 'ImportList' , and 'ExportList' is
determined by 'ImportList'. So it's fine to not track 'import type' for export list.
[1] 66cd8ec4c0/llvm/lib/LTO/LTO.cpp (L1815-L1819)
[2] 66cd8ec4c0/llvm/lib/LTO/LTO.cpp (L1783-L1794)
[3] 66cd8ec4c0/llvm/lib/LTO/LTO.cpp (L1494-L1496)
[4] b76100e220/llvm/lib/LTO/LTO.cpp (L194-L222)
SmallPtrSet.h and TimeProfiler.h are unused. CommandLine.h is only
needed for the UseNewDbgInfoFormat declare, which can be moved to the
places that need it.
This reverts commit e33db249b53fb70dce62db3ebd82d42239bd1d9d.
The change from *set to *map increases memory usage, and caused indexing
OOM in some applications. Need to profile offline to bring the memory
usage down.
For distributed ThinLTO, the LTO indexing step generates combined
summary for each module, and postlink pipeline reads the combined
summary which stores the information for link-time optimization.
This patch populates the 'import type' of a summary in bitcode, and
updates bitcode reader to parse the bit correctly.
Add 'sort' here since it's helpful when container type
changes (for example, https://github.com/llvm/llvm-project/pull/88024
wants to change container type from `unordered_set` to `DenseMap)
@MaskRay points out `std::` doesn't randomize the iteration order of
`unordered_{set,map}`, and the iteration order for single build is
deterministic.
We don't need to explicitly create these instances of ArrayRef because
Hasher::update takes ArrayRef, and ArrayRef can be implicitly
constructed from C arrays.
Directly load all bitcode into the new debug info format in `llvm-lto`
and `llvm-lto2`. This means that new-mode bitcode no longer round-trips
back to old-mode after parsing, and that old-mode bitcode gets
auto-upgraded to new-mode debug info (which is the current in-memory
default in LLVM).
The convention is for such MC-specific options to reside in
MCTargetOptions. However, CompressDebugSections/RelaxELFRelocations do
not follow the convention: `CompressDebugSections` is defined in both
TargetOptions and MCAsmInfo and there is forwarding complexity.
Move the option to MCTargetOptions and hereby simplify the code. Rename
the misleading RelaxELFRelocations to X86RelaxRelocations. llvm-mc
-relax-relocations and llc -x86-relax-relocations can now be unified.
The base class llvm::ThreadPoolInterface will be renamed
llvm::ThreadPool in a subsequent commit.
This is a breaking change: clients who use to create a ThreadPool must
now create a DefaultThreadPool instead.
The GlobalResolutions map was found to contribute ~9% of the peak
memory of a large thin link. However, we are essentially done with it
when we are about to compute cross module imports, which itself adds to
the peak memory due to the import and export lists (there is one use
just after importing but it can easily be moved before importing).
Move the last use up above importing, and free the GlobalResolutions
map after that (and before importing). To help guard against future
inadvertent use after it has been released, change it to a
std::optional.
Commit b963c0b658cc54b370832df4f5a3d63fd69da334 fixed LTO compilation of
cases where one translation unit is calling a function with the
dllimport attribute, and another translation unit provides this function
locally within the same linked module (i.e. not actually dllimported);
see https://github.com/llvm/llvm-project/issues/37453 or
https://bugs.llvm.org/show_bug.cgi?id=38105 for full context.
This was fixed by aliasing their GlobalResolution structs, for the
`__imp_` prefixed and non prefixed symbols.
I believe this fix to be wrong.
This patch reverts that fix, and fixes the same issue differently,
within LLD instead.
The fix assumed that one can treat the `__imp_` prefixed and unprefixed
symbols as equal, referencing SVN r240620
(d766653534e0cff702e42a43b44d3057b6094fea). However that referenced
commit had mistaken how this logic works, which was corrected later in
SVN r240622 (88e0f9206b4dccb56dee931adab08f89ff80525a); those symbols
aren't direct aliases for each other - but if there's a need for the
`__imp_` prefixed one and the other one exists, the `__imp_` prefixed
one is created, as a pointer to the other one.
However this fix only works if both translation units are compiled as
LTO; if the caller is compiled as a regular object file and the callee
is compiled as LTO, the fix fails, as the LTO compilation doesn't know
that the unprefixed symbol is needed.
The only level that knows of the potential relationship between the
`__imp_` prefixed and unprefixed symbol, across regular and bitcode
object files, is LLD itself.
Therefore, revert the original fix from
b963c0b658cc54b370832df4f5a3d63fd69da334, and fix the issue differently
- when concluding that we can fulfill an undefined symbol starting with
`__imp_`, mark the corresponding non prefixed symbol as used in a
regular object for the LTO compilation, to make sure that this non
prefixed symbol exists after the LTO compilation, to let LLD do the
fixup of the local import.
Extend the testcase to test a regular object file calling an LTO object
file, which previously failed.
This change also fixes another issue; an object file can provide both
unprefixed and prefixed versions of the same symbol, like this:
void importedFunc(void) {
}
void (*__imp_importedFunc)(void) = importedFunc;
That allows the function to be called both with and without dllimport
markings. (The concept of automatically resolving a reference to
`__imp_func` to a locally defined `func` only is done in MSVC style
linkers, but not in GNU ld, therefore MinGW mode code often uses this
construct.)
Previously, the aliasing of global resolutions at the LTO level would
trigger a failed assert with "Multiple prevailing defs are not allowed"
for this case, as both `importedFunc` and `__imp_importedFunc` could be
prevailing. Add a case to the existing LLD test case lto-imp-prefix.ll
to test this as well.
This change (together with previous change in
3ab6209a3f93bdbeec8e9b9fcc00a9a4980915ff) completes LLD to work with
mingw-w64-crt files (the base glue code for a mingw-w64 toolchain) built
with LTO.
Discussion about this approach: https://discourse.llvm.org/t/rfc-safer-whole-program-class-hierarchy-analysis/65144/18
When enabling WPD in an environment where native binaries are present, types we want to optimize can be derived from inside these native files and devirtualizing them can lead to correctness issues. RTTI can be used as a way to determine all such types in native files and exclude them from WPD providing a safe checked way to enable WPD.
The approach is:
1. In the linker, identify if RTTI is available for all native types. If not, under `--lto-validate-all-vtables-have-type-infos` `--lto-whole-program-visibility` is automatically disabled. This is done by examining all .symtab symbols in object files and .dynsym symbols in DSOs for vtable (_ZTV) and typeinfo (_ZTI) symbols and ensuring there's always a match for every vtable symbol.
2. During thinlink, if `--lto-validate-all-vtables-have-type-infos` is set and RTTI is available for all native types, identify all typename (_ZTS) symbols via their corresponding typeinfo (_ZTI) symbols that are used natively or outside of our summary and exclude them from WPD.
Testing:
ninja check-all
large Meta service that uses boost, glog and libstdc++.so runs successfully with WPD via --lto-whole-program-visibility. Previously, native types in boost caused incorrect devirtualization that led to crashes.
Reviewed By: MaskRay, tejohnson
Differential Revision: https://reviews.llvm.org/D155659
This will make it easy for callers to see issues with and fix up calls
to createTargetMachine after a future change to the params of
TargetMachine.
This matches other nearby enums.
For downstream users, this should be a fairly straightforward
replacement,
e.g. s/CodeGenOpt::Aggressive/CodeGenOptLevel::Aggressive
or s/CGFT_/CodeGenFileType::
The module paths string table mapped to both an id sequentially assigned
during LTO linking, and the module hash. The former is leftover from
before the module hash was added for caching and subsequently replaced
use of the module id when renaming promoted symbols (to avoid affects
due to link order changes). The sequentially assigned module id was not
removed, however, as it was still a convenience when serializing to/from
bitcode and assembly.
This patch removes the module id from this table, since it isn't
strictly needed and can lead to confusion on when it is appropriate to
use (e.g. see fix in D156525). It also takes a (likely not significant)
amount of overhead. Where an integer module id is needed (e.g. bitcode
writing), one is assigned on the fly.
There are a couple of test changes since the paths are now sorted
alphanumerically when assigning ids on the fly during assembly writing,
in order to ensure deterministic behavior.
Differential Revision: https://reviews.llvm.org/D156730
D151965 removed incorrect internalization for {linkonce,weak}{,_odr} when
the prevailing copy is in native code. The multiple conditions are based
on negative conditions, which can be simplified to be based on positive cases:
* an object with an external linkage (must be prevailing) can be internalized
* a prevailing object with a {linkonce,weak}{,_odr} or common linkage can be internalized.
Further, the lengthy comment is a bit misleading, as it doesn't say that
objects with an external/linkonce/weak linkage can be internalized.
Clarify it.
Reviewed By: tejohnson
Differential Revision: https://reviews.llvm.org/D158949
