Fixes issue where if a line section doesn't start with a line number
then the addresses at the beginning of the section don't have line numbers.
For example, for a line section like this
```
0001:00000010-00000014, line/column/addr entries = 1
7 00000013 !
```
a line number wouldn't be found for addresses from 10 to 12.
This matches behavior when using the DIA SDK.
Differential Revision: https://reviews.llvm.org/D93306
The existing code handles this correctly and I checked that the code
in NativeInlineSiteSymbol also handles this correctly, but it was
wrong in the NativeFunctionSymbol code.
Differential Revision: https://reviews.llvm.org/D92134
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
-Use the actual sect/offset to keep track of symbols in the cache so they don't get created multiple times with different addresses.
-Remove getSymTag from PDBFunctionSymbol/PDBPublicSymbol because it's already implemented in the base class
-Merge the symbolizer test files for DIA and native, since the tests are the same.
-Implement getCompilandId for NativeLineNumber
Reviewed By: amccarth
Differential Revision: https://reviews.llvm.org/D84208
When building in Debug on Windows-MSVC after b7402edce3, a lot of tests were failing because we were dereferencing an element past the end of HashRecords. This happened towards the end of the table, in unused slots.
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
Summary: This implements searching for function symbols and public symbols by address.
More specifically,
-Implements NativeSession::findSymbolByAddress for function symbols and
public symbols. I think data symbols are also searched for, but isn't
implemented in this patch.
-Adds classes for NativeFunctionSymbol and NativePublicSymbol
-Adds a '-use-native-pdb-reader' option to llvm-symbolizer, for testing
purposes.
Reviewers: rnk, amccarth, labath
Subscribers: mgorny, hiraditya, MaskRay, rupprecht, llvm-commits
Tags: #llvm
Differential Revision: https://reviews.llvm.org/D79269
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
The number of public symbols is very large, and each deserialization
does a few heap allocations. The public symbols are serialized by the
linker, so we can assume they have the expected layout and use it
directly.
Saves O(#publics) temporary heap allocations and shrinks some data
structures.
This accounts for a large portion of the memory allocations in LLD.
This DebugSubsectionRecordBuilder object can be stored directly in
C13Builders, it mostly wraps other subsections.
Remove the container kind field from the object. It is always the same
for all elements in the vector, and we can pass it in during writing.
the tests pass on Linux.
Summary:
This change implements readFromExe, and calculating VA and RVA, which
are some of the functionalities that will be used for native PDB reading
for llvm symbolizer.
bug: https://bugs.llvm.org/show_bug.cgi?id=41795
Summary:
This change implements readFromExe, and calculating VA and RVA, which
are some of the functionalities that will be used for native PDB reading
for llvm symbolizer.
bug: https://bugs.llvm.org/show_bug.cgi?id=41795
Reviewers: hans, amccarth, rnk
Subscribers: hiraditya, llvm-commits
Tags: #llvm
Differential Revision: https://reviews.llvm.org/D78128
When emitting PDBs, the TypeStreamMerger class is used to merge .debug$T records from the input .OBJ files into the output .PDB stream.
Records in .OBJs are not required to be aligned on 4-bytes, and "The Netwide Assembler 2.14" generates non-aligned records.
When compiling with -DLLVM_ENABLE_ASSERTIONS=ON, an assert was triggered in MergingTypeTableBuilder when non-ghash merging was used.
With ghash merging there was no assert.
As a result, LLD could potentially generate a non-aligned TPI stream.
We now align records on 4-bytes when record indices are remapped, in TypeStreamMerger::remapIndices().
Differential Revision: https://reviews.llvm.org/D75081
This is how it should've been and brings it more in line with
std::string_view. There should be no functional change here.
This is mostly mechanical from a custom clang-tidy check, with a lot of
manual fixups. It uncovers a lot of minor inefficiencies.
This doesn't actually modify StringRef yet, I'll do that in a follow-up.
Currently injected-sources-native.test fails with "Expected<T>
value was in success state.
(Note: Expected<T> values in success mode must still be checked
prior to being destroyed)"
when llvm is compiled with LLVM_ENABLE_ABI_BREAKING_CHECKS in Release.
The problem is that getStringForID returns Expected<StringRef>
and Expected value must always be checked, even if it is in success state.
Checking with assert only helps in Debug and is wrong.
Differential revision: https://reviews.llvm.org/D69251
llvm-svn: 375492
David added the JamCRC implementation in r246590. More recently, Eugene
added a CRC-32 implementation in r357901, which falls back to zlib's
crc32 function if present.
These checksums are essentially the same, so having multiple
implementations seems unnecessary. This replaces the CRC-32
implementation with the simpler one from JamCRC, and implements the
JamCRC interface in terms of CRC-32 since this means it can use zlib's
implementation when available, saving a few bytes and potentially making
it faster.
JamCRC took an ArrayRef<char> argument, and CRC-32 took a StringRef.
This patch changes it to ArrayRef<uint8_t> which I think is the best
choice, and simplifies a few of the callers nicely.
Differential revision: https://reviews.llvm.org/D68570
llvm-svn: 374148
Now that we've moved to C++14, we no longer need the llvm::make_unique
implementation from STLExtras.h. This patch is a mechanical replacement
of (hopefully) all the llvm::make_unique instances across the monorepo.
llvm-svn: 369013
- getCompression() used to return a PDB_SourceCompression even though
the docs for IDiaInjectedSource are explicit about the return value
being compiler-dependent. Return an uint32_t instead, and make the
printing code handle unknown values better by printing "Unknown" and
the int value instead of not printing any compression.
- Print compressed contents as hex dump, not as string.
- Add compression type "DotNet", which is used (at least) by csc.exe,
the C# compiler. Also add a lengthy comment describing the stream
contents (derived from looking at the raw hex contents long enough
to see the GUIDs, which led me to the roslyn and mono implementations
for handling this).
- The native injected source dumper was dumping the contents of the
whole data stream -- but csc.exe writes a stream that's padded with
zero bytes to the next 512 boundary, and the dia api doesn't display
those padding bytes. So make NativeInjectedSource::getCode() do the
same thing.
Differential Revision: https://reviews.llvm.org/D64879
llvm-svn: 366386
`pretty -native -injected-sources -injected-source-content` works with
this patch, and produces identical output to the dia version.
Differential Revision: https://reviews.llvm.org/D64428
llvm-svn: 366236
The construction was explained in
https://reviews.llvm.org/D44810?id=139526#inline-391999 but reading the code
shouldn't require hunting down old reviews to understand it.
The precomputed list was missing an entry for the empty list case, and
one entry at the very end. (The current last entry is the last one where
3 * BucketCount fits in a signed int, but the reference implementation
uses unsigneds as far as I can tell, so there's room for one more entry.)
No behavior change for inputs seen in practice.
Differential Revision: https://reviews.llvm.org/D64738
llvm-svn: 366107
The traits object is only used by a few methods. Deserializing a hash
table and walking it is possible without the traits object, so it
shouldn't be required to build a dummy object for that use case.
The TraitsT object used to be a function template parameter before
r327647, this restores it to that state.
This makes it clear that the traits object isn't needed at all in 1 of
the current 3 uses of HashTable (and I am going to add another use that
doesn't need it), and that the default PdbHashTraits isn't used outside
of tests.
While here, also re-enable 3 checks in the test that were commented out
(which requires making HashTableInternals templated and giving FooBar
an operator==).
No intended behavior change.
Differential Revision: https://reviews.llvm.org/D64640
llvm-svn: 365974
All callers had a PDBFile object at hand, so call
Pdb.createIndexedStream() instead, which pre-populates all the arguments
(and returns nullptr for kInvalidStreamIndex).
Also change safelyCreateIndexedStream() to only take the string index,
and update callers. Make the method public and call it in two places
that manually did the bounds checking before.
No intended behavior change.
Differential Revision: https://reviews.llvm.org/D64633
llvm-svn: 365936
Summary: Deduplicate S_CONSTANTS when linking, if they have the same value.
Reviewers: rnk
Subscribers: hiraditya, llvm-commits
Tags: #llvm
Differential Revision: https://reviews.llvm.org/D63151
llvm-svn: 363089
lld-link used to write PDB files that DIA couldn't recover natvis
files from if:
- The global strings table was > 64kiB
- There were at least 3 natvis files
The cause was that the hash function for the /src/headerblock stream
was incorrect: It needs to be truncated to 16 bit.
If the global strings table was <= 64kiB, truncating to 16 bit is a
no-op, so this wasn't needed for small programs.
If there are only 1 or 2 natvis files, then the growth strategy in
HashTable::grow() would mean the hash table would have 2 buckets (for 1
natvis file) or 4 buckets (for 4 natvis files), and since the hash
function is used modulo number of buckets, and since 2 and 4 divide
0x10000, the missing `% 0x10000` is a no-op there too. For 3 natvis
files, the hash table grows to 6 buckets, which has a factor that's not
common with 0x10000 and the difference starts to matter.
Fixes PR41626.
Differential Revision: https://reviews.llvm.org/D61277
llvm-svn: 359515
Summary:
Now CVType and CVSymbol are effectively type-safe wrappers around
ArrayRef<uint8_t>. Make the kind() accessor load it from the
RecordPrefix, which is the same for types and symbols.
Reviewers: zturner, aganea
Subscribers: hiraditya, llvm-commits
Tags: #llvm
Differential Revision: https://reviews.llvm.org/D60018
llvm-svn: 357658
This avoids allocating a few KB of heap memory on startup, and instead
allocates these maps lazily. I noticed this while profiling LLD.
llvm-svn: 357192
Before, empty debug streams were written as 8 bytes (4 bytes signature + 4 bytes for the GlobalRefs count).
With this patch, unused empty streams aren't emitted anymore. Modules now encode 65535 as an 'unused stream' value, by convention.
Also fix the * Linker * contrib section which wasn't correctly emitted previously.
Differential Revision: https://reviews.llvm.org/D59502
llvm-svn: 356395
When type streams with forward references were merged using GHashes, cycles
were introduced in the debug info. This was caused by
GlobalTypeTableBuilder::insertRecordAs() not inserting the record on the second
pass, thus yielding an empty ArrayRef at that record slot. Later on, upon PDB
emission, TpiStreamBuilder::commit() would skip that empty record, thus
offseting all indices that came after in the stream.
This solution comes in two steps:
1. Fix the hash calculation, by doing a multiple-step resolution, iff there are
forward references in the input stream.
2. Fix merge by resolving with multiple passes, therefore moving records with
forward references at the end of the stream.
This patch also adds support for llvm-readoj --codeview-ghash.
Finally, fix dumpCodeViewMergedTypes() which previously could reference deleted
memory.
Fixes PR40221
Differential Revision: https://reviews.llvm.org/D57790
llvm-svn: 353412
Summary:
This patch fixes access to fpo streams in native pdb from DbiStream and makes
code consistent with DbiStreamBuilder.
Patch By: leonid.mashinskiy
Reviewers: zturner, aleksandr.urakov
Reviewed By: zturner
Subscribers: llvm-commits
Differential Revision: https://reviews.llvm.org/D56725
llvm-svn: 352615
PDBs contain several serialized hash tables. In the microsoft-pdb
repo published to support LLVM implementing PDB support, the
provided initializes the bucket count for the TPI and IPI streams
to the maximum size. This occurs in tpi.cpp L33 and tpi.cpp L398.
In the LLVM code for generating PDBs, these streams are created with
minimum number of buckets. This difference makes LLVM generated
PDBs slower for when used for debugging.
Patch by C.J. Hebert
Differential Revision: https://reviews.llvm.org/D56942
llvm-svn: 352117
to reflect the new license.
We understand that people may be surprised that we're moving the header
entirely to discuss the new license. We checked this carefully with the
Foundation's lawyer and we believe this is the correct approach.
Essentially, all code in the project is now made available by the LLVM
project under our new license, so you will see that the license headers
include that license only. Some of our contributors have contributed
code under our old license, and accordingly, we have retained a copy of
our old license notice in the top-level files in each project and
repository.
llvm-svn: 351636
VarStreamArray was built on the assumption that it is backed by a
StreamRef, and offset 0 of that StreamRef is the first byte of the first
record in the array.
This is a logical and intuitive assumption, but unfortunately we have
use cases where it doesn't hold. Specifically, a PDB module's symbol
stream is prefixed by 4 bytes containing a magic value, and the first
byte of record data in the array is actually at offset 4 of this byte
sequence.
Previously, we would just truncate the first 4 bytes and then construct
the VarStreamArray with the resulting StreamRef, so that offset 0 of the
underlying stream did correspond to the first byte of the first record,
but this is problematic, because symbol records reference other symbol
records by the absolute offset including that initial magic 4 bytes. So
if another record wants to refer to the first record in the array, it
would say "the record at offset 4".
This led to extremely confusing hacks and semantics in loading code, and
after spending 30 minutes trying to get some math right and failing, I
decided to fix this in the underlying implementation of VarStreamArray.
Now, we can say that a stream is skewed by a particular amount. This
way, when we access a record by absolute offset, we can use the same
values that the records themselves contain, instead of having to do
fixups.
Differential Revision: https://reviews.llvm.org/D55344
llvm-svn: 348499
Previously these were dropped. We now understand them sufficiently
well to start emitting them. From the debugger's perspective, this
now enables us to have debug info about typedefs (both global and
function-locally scoped)
Differential Revision: https://reviews.llvm.org/D55228
llvm-svn: 348306
Part of the patch to not build the hash map eagerly was omitted
due to a merge conflict. Add it back, which should fix the failing
tests.
llvm-svn: 348166
Summary:
This speeds up linking clang.exe/pdb with /DEBUG:GHASH by 31%, from
12.9s to 9.8s.
Symbol records are typically small (16.7 bytes on average), but we
processed them one at a time. CVSymbol is a relatively "large" type. It
wraps an ArrayRef<uint8_t> with a kind an optional 32-bit hash, which we
don't need. Before this change, each DbiModuleDescriptorBuilder would
maintain an array of CVSymbols, and would write them individually with a
BinaryItemStream.
With this change, we now add symbols that happen to appear contiguously
in bulk. For each .debug$S section (roughly one per function), we
allocate two copies, one for relocation, and one for realignment
purposes. For runs of symbols that go in the module stream, which is
most symbols, we now add them as a single ArrayRef<uint8_t>, so the
vector DbiModuleDescriptorBuilder is roughly linear in the number of
.debug$S sections (O(# funcs)) instead of the number of symbol records
(very large).
Some stats on symbol sizes for the curious:
PDB size: 507M
sym bytes: 316,508,016
sym count: 18,954,971
sym byte avg: 16.7
As future work, we may be able to skip copying symbol records in the
linker for realignment purposes if we make LLVM write them aligned into
the object file. We need to double check that such symbol records are
still compatible with link.exe, but if so, it's definitely worth doing,
since my profile shows we spend 500ms in memcpy in the symbol merging
code. We could potentially cut that in half by saving a copy.
Alternatively, we could apply the relocations *after* we iterate the
symbols. This would require some careful re-engineering of the
relocation processing code, though.
Reviewers: zturner, aganea, ruiu
Subscribers: hiraditya, llvm-commits
Differential Revision: https://reviews.llvm.org/D54554
llvm-svn: 347687
In a previous patch, we pre-processed the TPI stream in order to build
the reverse mapping from nested type -> parent type so that we could
accurately reconstruct a DeclContext hierarchy.
However, there were some issues. An LF_NESTTYPE record is really just a
typedef, so although it happens to be used to indicate the name of the
nested type and referring to the global record which defines the type,
it is also used for every other kind of nested typedef. When we rebuild
the DeclContext hierarchy, we want it to be as accurate as possible,
which means that if we have something like:
struct A {
struct B {};
using C = B;
};
We don't want to create two CXXRecordDecls in the AST each with the
exact same definition. We just want to create one for B and then
define C as an alias to B. Previously, however, it would not be able
to distinguish between the two cases and it would treat A::B and
A::C as being two classes each with separate definitions. We address
the first half of improving the pre-processing logic so that only
actual definitions are treated this way.
Later, in a followup patch, we can handle the case of nested
typedefs since we're already going to be enumerating the field list
anyway and this patch introduces the general framework for
distinguishing between the two cases.
Differential Revision: https://reviews.llvm.org/D54357
llvm-svn: 346786
This is mostly some cleanup done in the process of implementing
some basic support for types. I tried to split up the patch a
bit to get some of the NFC portion of the patch out into a separate
commit, and this is the result of that. It moves some code around,
deletes some spurious namespace qualifications, removes some
unnecessary header includes, forward declarations, etc.
llvm-svn: 344913
Summary:
This patch just extends the `IPDBSession` interface to allow retrieving
of frame data through it, and adds an implementation over DIA. It is needed
for an implementation (for now with DIA) of the conversion from FPO programs
to DWARF expressions mentioned in D53086.
Reviewers: zturner, asmith, rnk
Reviewed By: asmith
Subscribers: mgorny, aprantl, JDevlieghere, llvm-commits
Differential Revision: https://reviews.llvm.org/D53324
llvm-svn: 344886
When we're on the last bucket the computation is tricky.
We were failing when the last bucket contained multiple
matches. Added a new test for this.
llvm-svn: 344081
We changed an ArrayRef<uint8_t> to an ArrayRef<uint32_t>, but
it needs to be an ArrayRef<support::ulittle32_t>.
We also change ArrayRef<> to FixedStreamArray<>. Technically
an ArrayRef<> will work, but it can cause a copy in the underlying
implementation if the memory is not contiguous, and there's no
reason not to use a FixedStreamArray<>.
Thanks to nemanjai@ and thakis@ for helping me track this down
and confirm the fix.
llvm-svn: 344063
Fix the following warning when compiling with clang (caused by commit
rL343951):
GlobalsStream.cpp:61:33: warning: comparison of integers of different
signs: 'int' and 'uint32_t'
This also avoids double evaluation of `GlobalsTable.HashBuckets.size()`.
llvm-svn: 343957
The Globals table is a hash table keyed on symbol name, so
it's possible to lookup symbols by name in O(1) time. Add
a function to the globals stream to do this, and add an option
to llvm-pdbutil to exercise this, then use it to write some
tests to verify correctness.
llvm-svn: 343951
These work a little differently because they are actually in
the globals stream and are treated as symbol records, even though
DIA presents them as types. So this also adds the necessary
infrastructure to cache records that live somewhere other than
the TPI stream as well.
llvm-svn: 343507
We didn't properly detect when a pointer was a member
pointer, and when that was the case we were not
properly returning class parent info. This caused
member pointers to render incorrectly in pretty mode.
However, we didn't even have pretty tests for pointers
in native mode, so those are also added now to ensure
this.
llvm-svn: 343393
This allows the native reader to find records of class/struct/
union type and dump them. This behavior is tested by using the
diadump subcommand against golden output produced by actual DIA
SDK on the same PDB file, and again using pretty -native to
confirm that we actually dump the classes. We don't find class
members or anything like that yet, for now it's just the class
itself.
llvm-svn: 342779
Some records point to an LF_CLASS, LF_UNION, LF_STRUCTURE, or LF_ENUM
which is a forward reference and doesn't contain complete debug
information. In these cases, we'd like to be able to quickly locate the
full record. The TPI stream stores an array of pre-computed record hash
values, one for each type record. If we pre-process this on startup, we
can build a mapping from hash value -> {list of possible matching type
indices}. Since hashes of full records are only based on the name and or
unique name and not the full record contents, we can then use forward
ref record to compute the hash of what *would* be the full record by
just hashing the name, use this to get the list of possible matches, and
iterate those looking for a match on name or unique name.
llvm-pdbutil is updated to resolve forward references for the purposes
of testing (plus it's just useful).
Differential Revision: https://reviews.llvm.org/D52283
llvm-svn: 342656
There were several issues with the previous implementation.
1) There were no tests.
2) We didn't support creating PDBSymbolTypePointer records for
builtin types since those aren't described by LF_POINTER
records.
3) We didn't support a wide enough variety of builtin types even
ignoring pointers.
This patch fixes all of these issues. In order to add tests,
it's helpful to be able to ignore the symbol index id hierarchy
because it makes the golden output from the DIA version not match
our output, so I've extended the dumper to disable dumping of id
fields.
llvm-svn: 342493
Previously we would dump the names of enum types, but not their
enumerator values. This adds support for enumerator values. In
doing so, we have to introduce a general purpose mechanism for
caching symbol indices of field list members. Unlike global
types, FieldList members do not have a TypeIndex. So instead,
we identify them by the pair {TypeIndexOfFieldList, IndexInFieldList}.
llvm-svn: 342415
Previously for cv-qualified types, we would just ignore them
and they would never get printed. Now we can enumerate them
and cache them like any other symbol type.
llvm-svn: 342414
Naively computing the hash after the PDB data has been generated is in practice
as fast as other approaches I tried. I also tried online-computing the hash as
parts of the PDB were written out (https://reviews.llvm.org/D51887; that's also
where all the measuring data is) and computing the hash in parallel
(https://reviews.llvm.org/D51957). This approach here is simplest, without
being slower.
Differential Revision: https://reviews.llvm.org/D51956
llvm-svn: 342333
Eventually we need to be able to support nested types, which don't
have an associated CVType record. To handle this, remove the
CVType from all of the record classes, and instead store the
deserialized record. Then move the deserialization up to the thing
that creates the type. This actually makes error handling better
anyway as we can return an invalid symbol instead of asserting false.
llvm-svn: 342284
r342003 added support for emitting FPO data from the
DEBUG_S_FRAMEDATA subsection of the .debug$S section to the PDB
file. However, that is not the end of the story. FPO can end
up in two different destinations in a PDB, each corresponding to
a different FPO data source.
The case handled by r342003 involves copying data from the
DEBUG_S_FRAMEDATA subsection of the .debug$S section to the
"New FPO" stream in the PDB, which is then referred to by the
DBI stream. The case handled by this patch involves copying
records from the .debug$F section of an object file to the "FPO"
stream (or perhaps more aptly, the "Old FPO" stream) in the PDB
file, which is also referred to by the DBI stream.
The formats are largely similar, and the difference is mostly
only visible in masm generated object files, such as some of the
low-level CRT object files like memcpy. MASM doesn't appear to
support writing the DEBUG_S_FRAMEDATA subsection, and instead
just writes these records to the .debug$F section.
Although clang-cl does not emit a .debug$F section ever, lld still
needs to support it so we have good debugging for CRT functions.
Differential Revision: https://reviews.llvm.org/D51958
llvm-svn: 342080
Makes the produced pdbs more deterministic; before they'd contain 2 arbitary
bytes where this padding was.
Also reorder initialization to match the order of the fields in the struct (nfc)
llvm-svn: 341945
They were unintentionally calling DIA directly, which requires
Windows. We need to pass the -native flag, and this then required
fixing up one or two tests.
llvm-svn: 341731
Kazu Hirata