As they can be set independently after D88802, we can get rid of a bit
of extra code - simplifying the logic here before adding more
complication to it later.
Differential Revision: https://reviews.llvm.org/D88803
The MinGW driver has separate options for OS and subsystem version.
Having this available in lld-link allows the MinGW driver to both match
GNU ld better and simplifies the code for merging two (potentially
mismatching) arguments into one.
Differential Revision: https://reviews.llvm.org/D88802
Parse the components as decimal, instead of decuding the base from
the string. This avoids ambiguity if the second number contains leading
zeros, which previously were parsed as indicating an octal number.
MS link.exe doesn't support hexadecimal numbers in the version numbers,
neither in /version nor in /subsystem.
Differential Revision: https://reviews.llvm.org/D88801
This adds the following two new lines to /summary:
21351 Input OBJ files (expanded from all cmd-line inputs)
61 PDB type server dependencies
38 Precomp OBJ dependencies
1420669231 Input type records <<<<
78665073382 Input type records bytes <<<<
8801393 Merged TPI records
3177158 Merged IPI records
59194 Output PDB strings
71576766 Global symbol records
25416935 Module symbol records
2103431 Public symbol records
Differential Revision: https://reviews.llvm.org/D88703
Before this patch /summary was crashing with some .PCH.OBJ files, because tpiMap[srcIdx++] was reading at the wrong location. When the TpiSource depends on a .PCH.OBJ file, the types should be offset by the previously merged PCH.OBJ set of indices.
Differential Revision: https://reviews.llvm.org/D88678
This patch adds support for creating Guard Address-Taken IAT Entry Tables (.giats$y sections) in object files, matching the behavior of MSVC. These contain lists of address-taken imported functions, which are used by the linker to create the final GIATS table.
Additionally, if any DLLs are delay-loaded, the linker must look through the .giats tables and add the respective load thunks of address-taken imports to the GFIDS table, as these are also valid call targets.
Reviewed By: rnk
Differential Revision: https://reviews.llvm.org/D87544
Stored Error objects have to be checked, even if they are success
values.
This reverts commit 8d250ac3cd.
Relands commit 49b3459930655d879b2dc190ff8fe11c38a8be5f..
Original commit message:
-----------------------------------------
This makes type merging much faster (-24% on chrome.dll) when multiple
threads are available, but it slightly increases the time to link (+10%)
when /threads:1 is passed. With only one more thread, the new type
merging is faster (-11%). The output PDB should be identical to what it
was before this change.
To give an idea, here is the /time output placed side by side:
BEFORE | AFTER
Input File Reading: 956 ms | 968 ms
Code Layout: 258 ms | 190 ms
Commit Output File: 6 ms | 7 ms
PDB Emission (Cumulative): 6691 ms | 4253 ms
Add Objects: 4341 ms | 2927 ms
Type Merging: 2814 ms | 1269 ms -55%!
Symbol Merging: 1509 ms | 1645 ms
Publics Stream Layout: 111 ms | 112 ms
TPI Stream Layout: 764 ms | 26 ms trivial
Commit to Disk: 1322 ms | 1036 ms -300ms
----------------------------------------- --------
Total Link Time: 8416 ms 5882 ms -30% overall
The main source of the additional overhead in the single-threaded case
is the need to iterate all .debug$T sections up front to check which
type records should go in the IPI stream. See fillIsItemIndexFromDebugT.
With changes to the .debug$H section, we could pre-calculate this info
and eliminate the need to do this walk up front. That should restore
single-threaded performance back to what it was before this change.
This change will cause LLD to be much more parallel than it used to, and
for users who do multiple links in parallel, it could regress
performance. However, when the user is only doing one link, it's a huge
improvement. In the future, we can use NT worker threads to avoid
oversaturating the machine with work, but for now, this is such an
improvement for the single-link use case that I think we should land
this as is.
Algorithm
----------
Before this change, we essentially used a
DenseMap<GloballyHashedType, TypeIndex> to check if a type has already
been seen, and if it hasn't been seen, insert it now and use the next
available type index for it in the destination type stream. DenseMap
does not support concurrent insertion, and even if it did, the linker
must be deterministic: it cannot produce different PDBs by using
different numbers of threads. The output type stream must be in the same
order regardless of the order of hash table insertions.
In order to create a hash table that supports concurrent insertion, the
table cells must be small enough that they can be updated atomically.
The algorithm I used for updating the table using linear probing is
described in this paper, "Concurrent Hash Tables: Fast and General(?)!":
https://dl.acm.org/doi/10.1145/3309206
The GHashCell in this change is essentially a pair of 32-bit integer
indices: <sourceIndex, typeIndex>. The sourceIndex is the index of the
TpiSource object, and it represents an input type stream. The typeIndex
is the index of the type in the stream. Together, we have something like
a ragged 2D array of ghashes, which can be looked up as:
tpiSources[tpiSrcIndex]->ghashes[typeIndex]
By using these side tables, we can omit the key data from the hash
table, and keep the table cell small. There is a cost to this: resolving
hash table collisions requires many more loads than simply looking at
the key in the same cache line as the insertion position. However, most
supported platforms should have a 64-bit CAS operation to update the
cell atomically.
To make the result of concurrent insertion deterministic, the cell
payloads must have a priority function. Defining one is pretty
straightforward: compare the two 32-bit numbers as a combined 64-bit
number. This means that types coming from inputs earlier on the command
line have a higher priority and are more likely to appear earlier in the
final PDB type stream than types from an input appearing later on the
link line.
After table insertion, the non-empty cells in the table can be copied
out of the main table and sorted by priority to determine the ordering
of the final type index stream. At this point, item and type records
must be separated, either by sorting or by splitting into two arrays,
and I chose sorting. This is why the GHashCell must contain the isItem
bit.
Once the final PDB TPI stream ordering is known, we need to compute a
mapping from source type index to PDB type index. To avoid starting over
from scratch and looking up every type again by its ghash, we save the
insertion position of every hash table insertion during the first
insertion phase. Because the table does not support rehashing, the
insertion position is stable. Using the array of insertion positions
indexed by source type index, we can replace the source type indices in
the ghash table cells with the PDB type indices.
Once the table cells have been updated to contain PDB type indices, the
mapping for each type source can be computed in parallel. Simply iterate
the list of cell positions and replace them with the PDB type index,
since the insertion positions are no longer needed.
Once we have a source to destination type index mapping for every type
source, there are no more data dependencies. We know which type records
are "unique" (not duplicates), and what their final type indices will
be. We can do the remapping in parallel, and accumulate type sizes and
type hashes in parallel by type source.
Lastly, TPI stream layout must be done serially. Accumulate all the type
records, sizes, and hashes, and add them to the PDB.
Differential Revision: https://reviews.llvm.org/D87805
This makes type merging much faster (-24% on chrome.dll) when multiple
threads are available, but it slightly increases the time to link (+10%)
when /threads:1 is passed. With only one more thread, the new type
merging is faster (-11%). The output PDB should be identical to what it
was before this change.
To give an idea, here is the /time output placed side by side:
BEFORE | AFTER
Input File Reading: 956 ms | 968 ms
Code Layout: 258 ms | 190 ms
Commit Output File: 6 ms | 7 ms
PDB Emission (Cumulative): 6691 ms | 4253 ms
Add Objects: 4341 ms | 2927 ms
Type Merging: 2814 ms | 1269 ms -55%!
Symbol Merging: 1509 ms | 1645 ms
Publics Stream Layout: 111 ms | 112 ms
TPI Stream Layout: 764 ms | 26 ms trivial
Commit to Disk: 1322 ms | 1036 ms -300ms
----------------------------------------- --------
Total Link Time: 8416 ms 5882 ms -30% overall
The main source of the additional overhead in the single-threaded case
is the need to iterate all .debug$T sections up front to check which
type records should go in the IPI stream. See fillIsItemIndexFromDebugT.
With changes to the .debug$H section, we could pre-calculate this info
and eliminate the need to do this walk up front. That should restore
single-threaded performance back to what it was before this change.
This change will cause LLD to be much more parallel than it used to, and
for users who do multiple links in parallel, it could regress
performance. However, when the user is only doing one link, it's a huge
improvement. In the future, we can use NT worker threads to avoid
oversaturating the machine with work, but for now, this is such an
improvement for the single-link use case that I think we should land
this as is.
Algorithm
----------
Before this change, we essentially used a
DenseMap<GloballyHashedType, TypeIndex> to check if a type has already
been seen, and if it hasn't been seen, insert it now and use the next
available type index for it in the destination type stream. DenseMap
does not support concurrent insertion, and even if it did, the linker
must be deterministic: it cannot produce different PDBs by using
different numbers of threads. The output type stream must be in the same
order regardless of the order of hash table insertions.
In order to create a hash table that supports concurrent insertion, the
table cells must be small enough that they can be updated atomically.
The algorithm I used for updating the table using linear probing is
described in this paper, "Concurrent Hash Tables: Fast and General(?)!":
https://dl.acm.org/doi/10.1145/3309206
The GHashCell in this change is essentially a pair of 32-bit integer
indices: <sourceIndex, typeIndex>. The sourceIndex is the index of the
TpiSource object, and it represents an input type stream. The typeIndex
is the index of the type in the stream. Together, we have something like
a ragged 2D array of ghashes, which can be looked up as:
tpiSources[tpiSrcIndex]->ghashes[typeIndex]
By using these side tables, we can omit the key data from the hash
table, and keep the table cell small. There is a cost to this: resolving
hash table collisions requires many more loads than simply looking at
the key in the same cache line as the insertion position. However, most
supported platforms should have a 64-bit CAS operation to update the
cell atomically.
To make the result of concurrent insertion deterministic, the cell
payloads must have a priority function. Defining one is pretty
straightforward: compare the two 32-bit numbers as a combined 64-bit
number. This means that types coming from inputs earlier on the command
line have a higher priority and are more likely to appear earlier in the
final PDB type stream than types from an input appearing later on the
link line.
After table insertion, the non-empty cells in the table can be copied
out of the main table and sorted by priority to determine the ordering
of the final type index stream. At this point, item and type records
must be separated, either by sorting or by splitting into two arrays,
and I chose sorting. This is why the GHashCell must contain the isItem
bit.
Once the final PDB TPI stream ordering is known, we need to compute a
mapping from source type index to PDB type index. To avoid starting over
from scratch and looking up every type again by its ghash, we save the
insertion position of every hash table insertion during the first
insertion phase. Because the table does not support rehashing, the
insertion position is stable. Using the array of insertion positions
indexed by source type index, we can replace the source type indices in
the ghash table cells with the PDB type indices.
Once the table cells have been updated to contain PDB type indices, the
mapping for each type source can be computed in parallel. Simply iterate
the list of cell positions and replace them with the PDB type index,
since the insertion positions are no longer needed.
Once we have a source to destination type index mapping for every type
source, there are no more data dependencies. We know which type records
are "unique" (not duplicates), and what their final type indices will
be. We can do the remapping in parallel, and accumulate type sizes and
type hashes in parallel by type source.
Lastly, TPI stream layout must be done serially. Accumulate all the type
records, sizes, and hashes, and add them to the PDB.
Differential Revision: https://reviews.llvm.org/D87805
Library users should not need to call errorHandler().reset() explicitly.
google/iree calls lld:🧝:link and without the patch some global
variables are not cleaned up in the next invocation.
In lit tests, we run each LLD invocation twice (LLD_IN_TEST=2), without shutting down the process in-between. This ensures a full cleanup is properly done between runs.
Only active for the COFF driver for now. Other drivers still use LLD_IN_TEST=1 which executes just one iteration with full cleanup, like before.
When the environment variable LLD_IN_TEST is unset, a shortcut is taken, only one iteration is executed, no cleanup for faster exit, like before.
A public API, lld::safeLldMain(), is also available when using LLD as a library.
Differential Revision: https://reviews.llvm.org/D70378
Extending the lifetime of these type index mappings does increase memory
usage (+2% in my case), but it decouples type merging from symbol
merging. This is a pre-requisite for two changes that I have in mind:
- parallel type merging: speeds up slow type merging
- defered symbol merging: avoid heap allocating (relocating) all symbols
This eliminates CVIndexMap and moves its data into TpiSource. The maps
are also split into a SmallVector and ArrayRef component, so that the
ipiMap can alias the tpiMap for /Z7 object files, and so that both maps
can simply alias the PDB type server maps for /Zi files.
Splitting TypeServerSource establishes that all input types to be merged
can be identified with two 32-bit indices:
- The index of the TpiSource object
- The type index of the record
This is useful, because this information can be stored in a single
64-bit atomic word to enable concurrent hashtable insertion.
One last change is that now all object files with debugChunks get a
TpiSource, even if they have no type info. This avoids some null checks
and special cases.
Differential Revision: https://reviews.llvm.org/D87736
Binutils generated sections seem to be padded to a multiple of 16 bytes,
but the aux section definition contains the original, unpadded section
length.
The size check used for IMAGE_COMDAT_SELECT_SAME_SIZE previously
only checked the size of the section itself. When checking the
currently processed object file against the previously chosen
comdat section, we easily have access to the aux section definition
of the currently processed section, but we have to iterate over the
symbols of the previously selected object file to find the section
definition of the previously picked section. (We don't want to
inflate SectionChunk to carry more data, for something that is only
needed in corner cases.) Only do this when the mingw flag is set.
This fixes statically linking clang-built C++ object files against
libstdc++ built with GCC, if the object files contain e.g. typeinfo.
Differential Revision: https://reviews.llvm.org/D86659
The global variable outputSections in the COFF writer was not
cleared between runs which caused successive calls to lld::coff::link
to generate invalid binaries. These binaries when loaded would result
in "invalid win32 applications" and/or "bad image" errors.
Differential Revision: https://reviews.llvm.org/D86401
This patch adds the missing information to the LF_BUILDINFO record, which allows for rebuilding a .CPP without any external dependency but the .OBJ itself (other than the compiler).
Some external tools that we are using (Recode, Live++) are extracting the information to reproduce a build without any knowledge of the build system. The LF_BUILDINFO stores a full path to the compiler, the PWD (CWD at program startup), a relative or absolute path to the TU, and the full CC1 command line. The command line needs to be freestanding (not depend on any environment variables). In the same way, MSVC doesn't store the provided command-line, but an expanded version (somehow their equivalent of CC1) which is also freestanding.
For more information see PR36198 and D43002.
Differential Revision: https://reviews.llvm.org/D80833
Previously this flag was just ignored. If set, set the
IMAGE_DLL_CHARACTERISTICS_NO_SEH bit, regardless of the normal safeSEH
machinery.
In mingw configurations, the safeSEH bit might not be set in e.g. object
files built from handwritten assembly, making it impossible to use the
normal safeseh flag. As mingw setups don't generally use SEH on 32 bit
x86 at all, it should be fine to set that flag bit though - hook up
the existing GNU ld flag for controlling that.
Differential Revision: https://reviews.llvm.org/D84701
For a weak symbol func in a comdat, the actual leader symbol ends up
named like .weak.func.default*. Likewise, for stdcall on i386, the symbol
may be named _func@4, while the section suffix only is "func", which the
previous implementation didn't handle.
This fixes unwinding through weak functions when using
-ffunction-sections in mingw environments.
Differential Revision: https://reviews.llvm.org/D84607
The "undefined symbol" error message from lld-link displays up to 3 references to that symbol, and the number of extra references not shown.
This patch removes the computation of the strings for those extra references.
It fixes a freeze of lld-link we accidentally encountered when activating asan on a large project, without linking with the asan library.
In that case, __asan_report_load8 was referenced more than 2 million times, causing the computation of that many display strings, of which only 3 were used.
Differential Revision: https://reviews.llvm.org/D83510
The `intrinsics_gen` target exists in the CMake exports since r309389
(see LLVMConfig.cmake.in), hence projects can depend on `intrinsics_gen`
even it they are built separately from LLVM.
Reviewed By: MaskRay, JDevlieghere
Differential Revision: https://reviews.llvm.org/D83454
Previously, lld would crash if the .pdata size was not an even multiple
of the expected .pdata entry size. This makes it error gracefully instead.
(We hit this in Chromium due to an assembler problem: https://crbug.com/1101577)
Differential revision: https://reviews.llvm.org/D83479
This patch adds some missing information to the LF_BUILDINFO which allows for rebuilding an .OBJ without any external dependency but the .OBJ itself (other than the compiler executable).
Some tools need this information to reproduce a build without any knowledge of the build system. The LF_BUILDINFO therefore stores a full path to the compiler, the PWD (which is the CWD at program startup), a relative or absolute path to the TU, and the full CC1 command line. The command line needs to be freestanding (not depend on any environment variable). In the same way, MSVC doesn't store the provided command-line, but an expanded version (somehow their equivalent of CC1) which is also freestanding.
For more information see PR36198 and D43002.
Differential Revision: https://reviews.llvm.org/D80833
This reduces peak memory on my test case from 1960.14MB to 1700.63MB
(-260MB, -13.2%) with no measurable impact on CPU time. I'm currently
working with a publics stream that is about 277MB. Before this change,
we would allocate 277MB of heap memory, serialize publics into them,
hold onto that heap memory, open the PDB, and commit into it. After
this change, we defer the serialization until commit time.
In the last change I made to public writing, I re-sorted the list of
publics multiple times in place to avoid allocating new temporary data
structures. Deferring serialization until later requires that we don't
reorder the publics. Instead of sorting the publics, I partially
construct the hash table data structures, store a publics index in them,
and then sort the hash table data structures. Later, I replace the index
with the symbol record offset.
This change also addresses a FIXME and moves the list of global and
public records from GSIHashStreamBuilder to GSIStreamBuilder. Now that
publics aren't being serialized, it makes even less sense to store them
as a list of CVSymbol records. The hash table used to deduplicate
globals is moved as well, since that is specific to globals, and not
publics.
Reviewed By: aganea, hans
Differential Revision: https://reviews.llvm.org/D81296
This patch adds some missing information to the LF_BUILDINFO which allows for rebuilding an .OBJ without any external dependency but the .OBJ itself (other than the compiler executable).
Some tools need this information to reproduce a build without any knowledge of the build system. The LF_BUILDINFO therefore stores a full path to the compiler, the PWD (which is the CWD at program startup), a relative or absolute path to the TU, and the full CC1 command line. The command line needs to be freestanding (not depend on any environment variable). In the same way, MSVC doesn't store the provided command-line, but an expanded version (somehow their equivalent of CC1) which is also freestanding.
For more information see PR36198 and D43002.
Differential Revision: https://reviews.llvm.org/D80833
First, do not reserve numSections in the Chunks array. In cases where
there are many non-prevailing sections, this will overallocate memory
which will not be used.
Second, free the memory for sparseChunks after initializeSymbols. After
that, it is never used.
This saves 50MB of 627MB for my use case without affecting performance.
The inlinees section contains references to the file checksum table. The
file checksum table in the PDB must have the same layout as the file
checksum table in the object file, so all the existing file id
references should stay valid.
Previously, we would do this:
for all inlined functions:
- lookup filename from checksum and string table
- make that filename absolute
- look up the new file id for that filename up in the new checksum
table
This lead to pdbMakeAbsolute and remove_dots ending up in the hot path.
We should only need to absolutify the source path once, not once every
time we process an inline function from that source file.
This speeds up linking chrome PGO stage 1 net_unittests.exe from 9.203s
to 8.500s (-7.6%). Looking just at time to process symbol records, it
goes from ~2000ms to ~1300ms, which is consistent with the overall
speedup of about 700ms. This will be less noticeable in debug builds,
which have fewer inlined functions records.
Summary:
This is a pre-requisite to parallelizing PDB symbol and type merging.
Currently this timer usage would not be thread safe.
Reviewers: aganea, MaskRay
Subscribers: jfb, llvm-commits
Tags: #llvm
Differential Revision: https://reviews.llvm.org/D80298
This paves the way to doing more things in parallel, and allows us to
order type sources in dependency order. PDBs and PCH objects have to be
loaded before object files which use them.
This is a rebase of the unapplied remaining changes in
https://reviews.llvm.org/D59226. I found it very challenging to rebase
this across the LLD variable name style change. I recall there was a
tool for that, but I didn't take the time to use it.
Reviewers: aganea, akhuang
Subscribers: llvm-commits
Tags: #llvm
Differential Revision: https://reviews.llvm.org/D79672
Allow disabling either the full auto import feature, or just
forbidding the cases that require runtime fixups.
As long as all auto imported variables are referenced from separate
.refptr$<name> sections, we can alias them on top of the IAT entries
and don't actually need any runtime fixups via pseudo relocations.
LLVM generates references to variables in .refptr stubs, if it
isn't known that the variable for sure is defined in the same object
module. Runtime pseudo relocs are needed if the addresses of auto
imported variables are used in constant initializers though.
Fixing up runtime pseudo relocations requires the use of
VirtualProtect (which is disallowed in WinStore/UWP apps) or
VirtualProtectFromApp. To allow any risk of ambiguity, allow
rejecting cases that would require this at the linker stage.
This adds support for the --disable-runtime-pseudo-reloc and
--disable-auto-import options in the MinGW driver (matching GNU ld.bfd)
with corresponding lld private options in the COFF driver.
Differential Revision: https://reviews.llvm.org/D78923
This fixes an accidental breakage of exporting symbols using def
files, when the symbol name contains a period, since commit
0ca06f7950, mixing up a symbol name containing a period with
the case of exporting a symbol as a forward to another dll.
Differential Revision: https://reviews.llvm.org/D79619
I noticed that std::error_code() does one-time initialization. Avoid
that overhead with Expected<T> and llvm::Error. Also, it is consistent
with the virtual interface and ELF, and generally cleaner.
Reviewed By: MaskRay
Differential Revision: https://reviews.llvm.org/D79643
Reduces time to link PGO instrumented net_unittets.exe by 11% (9.766s ->
8.672s, best of three). Reduces peak memory by 65.7MB (2142.71MB ->
2076.95MB).
Use a more compact struct, BulkPublic, for faster sorting. Sort in
parallel. Construct the hash buckets in parallel. Try to use one vector
to hold all the publics instead of copying them from one to another.
Allocate all the memory needed to serialize publics up front, and then
serialize them in place in parallel.
Reviewed By: aganea, hans
Differential Revision: https://reviews.llvm.org/D79467
Before this patch, the debug record S_GTHREAD32 which represents global thread_local symbols, was emitted by LLD into the respective module stream. This makes Visual Studio unable to display thread_local symbols in the debugger.
After this patch, S_GTHREAD32 is moved into the globals stream. This matches MSVC behavior.
Differential Revision: https://reviews.llvm.org/D79005
Essentially takes the lld/Common/Threads.h wrappers and moves them to
the llvm/Support/Paralle.h algorithm header.
The changes are:
- Remove policy parameter, since all clients use `par`.
- Rename the methods to `parallelSort` etc to match LLVM style, since
they are no longer C++17 pstl compatible.
- Move algorithms from llvm::parallel:: to llvm::, since they have
"parallel" in the name and are no longer overloads of the regular
algorithms.
- Add range overloads
- Use the sequential algorithm directly when 1 thread is requested
(skips task grouping)
- Fix the index type of parallelForEachN to size_t. Nobody in LLVM was
using any other parameter, and it made overload resolution hard for
for_each_n(par, 0, foo.size(), ...) because 0 is int, not size_t.
Remove Threads.h and update LLD for that.
This is a prerequisite for parallel public symbol processing in the PDB
library, which is in LLVM.
Reviewed By: MaskRay, aganea
Differential Revision: https://reviews.llvm.org/D79390
Summary:
That unless the user requested an output object (--lto-obj-path), the an
unused empty combined module is not emitted.
This changed is helpful for some target (ex. RISCV-V) which encoded the
ABI info in IR module flags (target-abi). Empty unused module has no ABI
info so the linker would get the linking error during merging
incompatible ABIs.
Reviewers: tejohnson, espindola, MaskRay
Subscribers: emaste, inglorion, arichardson, hiraditya, simoncook, MaskRay, steven_wu, dexonsmith, PkmX, dang, lenary, s.egerton, luismarques, llvm-commits
Tags: #llvm
Differential Revision: https://reviews.llvm.org/D78988
Heap profiling with ETW shows that LLD performs 4,053,721 heap
allocations over its lifetime, and ~800,000 of them come from
assocEquals. These vectors are created just to do a comparison, so fuse
the comparison into the loop and avoid the allocation.
ICF is overall a small portion of the time spent linking, and I did not
measure overall throughput improvements from this change above the noise
threshold. However, these show up in the heap profiler, and the work is
done, so we might as well land it if the code is clear enough.
Reviewed By: hans
Differential Revision: https://reviews.llvm.org/D79297
Martin Storsjö