This is a pretty classic optimization. Instead of processing symbol
records and copying them to temporary storage, do a first pass to
measure how large the module symbol stream will be, and then copy the
data into place in the PDB file. This requires defering relocation until
much later, which accounts for most of the complexity in this patch.
This patch avoids copying the contents of all live .debug$S sections
into heap memory, which is worth about 20% of private memory usage when
making PDBs. However, this is not an unmitigated performance win,
because it can be faster to read dense, temporary, heap data than it is
to iterate symbol records in object file backed memory a second time.
Results on release chrome.dll:
peak mem: 5164.89MB -> 4072.19MB (-1,092.7MB, -21.2%)
wall-j1: 0m30.844s -> 0m32.094s (slightly slower)
wall-j3: 0m20.968s -> 0m20.312s (slightly faster)
wall-j8: 0m19.062s -> 0m17.672s (meaningfully faster)
I gathered similar numbers for a debug, component build of content.dll
in Chrome, and the performance impact of this change was in the noise.
The memory usage reduction was visible and similar.
Because of the new parallelism in the PDB commit phase, more cores makes
the new approach faster. I'm assuming that most C++ developer machines
these days are at least quad core, so I think this is a win.
Differential Revision: https://reviews.llvm.org/D94267
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
llvm-symbolizer used to use the DIA SDK for symbolization on
Windows; this patch switches to using native symbolization, which was
implemented recently.
Users can still make the symbolizer use DIA by adding the `-dia` flag
in the LLVM_SYMBOLIZER_OPTS environment variable.
Differential Revision: https://reviews.llvm.org/D91814
This allows to reuse the RelocationResolver from the code
that doesn't want to deal with `RelocationRef` class.
I am going to use it in llvm-readobj. See the description
of D91530 for more details.
Differential revision: https://reviews.llvm.org/D91533
In the current state, if getFromHash(0) is called and there's no CU with
dwo_id=0, the lookup will stop at an empty slot, then the check
`Rows[H].getSignature() != S` won't cause the lookup to fail and return
a nullptr (as it should), because the empty slot has a 0 in the
signature field, and a pointer to the empty slot will be incorrectly
returned.
This patch fixes this by using the index field in the hash entry to
check for empty slots: signature = 0 can match a valid hash but
according to the spec the index for an occupied slot will always be
non-zero.
Differential Revision: https://reviews.llvm.org/D91670
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
When compiling for Windows on Arm the amd64 debug interfce from the Visual
Studio SDK is used as the cmake currently only distinguishes between x86 and
amd64 by checking the pointer size. Instead we can get the target
architecture for the compilier and check that to distinguish between
architectures.
We used to only emit static const data members in CodeView as
S_CONSTANTS when they were used; this patch makes it so they are always emitted.
This changes CodeViewDebug.cpp to find the static const members from the
class debug info instead of creating DIGlobalVariables in the IR
whenever a static const data member is used.
Bug: https://bugs.llvm.org/show_bug.cgi?id=47580
Differential Revision: https://reviews.llvm.org/D89072
This reverts commit 504615353f31136dd6bf7a971b6c236fd70582be.
We used to only emit static const data members in CodeView as
S_CONSTANTS when they were used; this patch makes it so they are always emitted.
I changed CodeViewDebug.cpp to find the static const members from the
class debug info instead of creating DIGlobalVariables in the IR
whenever a static const data member is used.
Bug: https://bugs.llvm.org/show_bug.cgi?id=47580
Differential Revision: https://reviews.llvm.org/D89072
Seems users have enough different uses of the symbolizer where they
might have unknown binaries and offsets such that "best effort" behavior
is all that's expected of llvm-symbolizer - so even erroring on unknown
executables and out of bounds offsets might not be suitable.
This reverts commit 1de0199748ef2a20cd146c100ea1b8e6726c4767.
This reverts commit a7b209a6d40d77b43a38664b1fe64513587f24c6.
This reverts commit 338dd138ea4a70b52ab48e0c8aa38ec152b3569a.
Create the LLVM / CodeView register mappings for the 32-bit ARM Window targets.
Reviewed By: compnerd
Differential Revision: https://reviews.llvm.org/D89622
There's no way to know whether there's a loclist contribution to parse
if there's no loclistx encoding - and if there is one, there's no need
to walk back from the loclist_base (or, uin the case of
info.dwo/loclist.dwo - starting at 0 in the contribution) to parse the
header, instead rely on the DWARF32/64 and address size in the CU
that's already available.
This would come up in split DWARF (non-split wouldn't try to read a
loclist header in the absence of a loclist_base) when one unit had
location lists and another does not (because the loclists.dwo section
would be non-empty in that case - in the case where it's empty the
parsing would silently skip).
Simplify the testing a bit, rather than needing a whole dwp, etc - by
creating a malformed loclists.dwo section (and use single file Split
DWARF) that would trip up any attempt to parse it - but no attempt
should be made.
Register context information was already being passed into the DWARFDebugFrame code that dumps unwind information but it wasn't being used. This change adds the ability to dump registers names of a valid MC register context was passed in and if it knows about the register. Updated the tests to use the newly returned register names.
Differential Revision: https://reviews.llvm.org/D88767
It's not possible to do this in complete generality - a CU using a
sec_offset DW_AT_ranges has no way of knowing where its rnglists
contribution starts, so should not attempt to parse any full rnglist
table/header to do so. And even using FORM_rnglistx there's no need to
parse the header - the offset can be computed using the CU's DWARF
format (32 or 64) to compute offset entry sizes, and then the list
parsed at that offset without ever trying to find a rnglist contribution
header immediately prior to the rnglists_base.
Stored Error objects have to be checked, even if they are success
values.
This reverts commit 8d250ac3cd48d0f17f9314685a85e77895c05351.
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
Since DWARFv5 places TUs in debug_info, some of DWARFContext's APIs have
become a bit erroneous, including TUs in the CU list by accident.
Correct that by providing compile_units (& dwo_compile_units) that
filter out the type units from the debug_info units.
Differential Revision: https://reviews.llvm.org/D87935
Flag DIEs that have DW_CHILDREN_yes set in their abbreviation but don't
actually have any children.
rdar://59809554
Differential revision: https://reviews.llvm.org/D88048
Most clients only need CVType and CVSymbol, not structs for every type
and symbol. Move CVSymbol and CVType to CVRecord.h to accomplish this.
Update some of the common headers that need CVSymbol and CVType to use
the new location.
When concatenating directory with filename in getFilenameByIndex, we
might end up with a path that contains extra dots. For example, if the
input is /path and ./example, we would return /path/./example. Run
sys::path::remove_dots on the output to eliminate unnecessary dots.
Differential Revision: https://reviews.llvm.org/D87657
Since a function might have portions of its code coming from multiple
different files, "start line" is ambiguous (it can't just be resolved
relative to the file/line specified). Add start file to disambiguate it.
When llvm-dwarfdump encounters no null terminated strings, we should
warn user about it rather than ignore it and print nothing.
Before this patch, when llvm-dwarfdump dumps a .debug_str section whose
content is "abc", it prints:
```
.debug_str contents:
```
After this patch:
```
.debug_str contents:
warning: no null terminated string at offset 0x0
```
Reviewed By: jhenderson, MaskRay
Differential Revision: https://reviews.llvm.org/D86998
This patch adds a helper function DumpStrSection to simplify codes.
Besides, nonprintable chars in debug_str and debug_str.dwo sections
are printed as escaped chars.
Reviewed By: jhenderson
Differential Revision: https://reviews.llvm.org/D86918
