**Note:** The register reading and writing depends on new register
flavor support in thread_get_state/thread_set_state in the kernel, which
will be first available in macOS 15.4.
The Apple M4 line of cores includes the Scalable Matrix Extension (SME)
feature. The M4s do not implement Scalable Vector Extension (SVE),
although the processor is in Streaming SVE Mode when the SME is being
used. The most obvious side effects of being in SSVE Mode are that (on
the M4 cores) NEON instructions cannot be used, and watchpoints may get
false positives, the address comparisons are done at a lowered
granularity.
When SSVE mode is enabled, the kernel will provide the Streaming Vector
Length register, which is a maximum of 64 bytes with the M4. Also
provided are SVCR (with bits indicating if SSVE mode and SME mode are
enabled), TPIDR2, SVL. Then the SVE registers Z0..31 (SVL bytes long),
P0..15 (SVL/8 bytes), the ZA matrix register (SVL*SVL bytes), and the M4
supports SME2, so the ZT0 register (64 bytes).
When SSVE/SME are disabled, none of these registers are provided by the
kernel - reads and writes of them will fail.
Unlike Linux, lldb cannot modify the SVL through a thread_set_state
call, or change the processor state's SSVE/SME status. There is also no
way for a process to request a lowered SVL size today, so the work that
David did to handle VL/SVL changing while stepping through a process is
not an issue on Darwin today. But debugserver should be providing
everything necessary so we can reuse all of David's work on resizing the
register contexts in lldb if it happens in the future. debugbserver
sends svl, svcr, and tpidr2 in the expedited registers when a thread
stops, if SSVE|SME mode are enabled (if the kernel allows it to read the
ARM_SME_STATE register set).
While the maximum SVL is 64 bytes on M4, the AArch64 maximum possible
SVL is 256; this would give us a 64k ZA register. If debugserver sized
all of its register contexts assuming the largest possible SVL, we could
easily use 2MB more memory for the register contexts of all threads in a
process -- and on iOS et al, processes must run within a small memory
allotment and this would push us over that.
Much of the work in debugserver was changing the arm64 register context
from being a static compile-time array of register sets, to being
initialized at runtime if debugserver is running on a machine with SME.
The ZA is only created to the machine's actual maximum SVL. The size of
the 32 SVE Z registers is less significant so I am statically allocating
those to the architecturally largest possible SVL value today.
Also, debugserver includes information about registers that share the
same part of the register file. e.g. S0 and D0 are the lower parts of
the NEON 128-bit V0 register. And when running on an SME machine, v0 is
the lower 128 bits of the SVE Z0 register. So the register maps used
when defining the VFP registers must differ depending on the
capabilities of the cpu at runtime.
I also changed register reading in debugserver, where formerly when
debugserver was asked to read a register, and the thread_get_state read
of that register failed, it would return all zero's. This is necessary
when constructing a `g` packet that gets all registers - because there
is no separation between register bytes, the offsets are fixed. But when
we are asking for a single register (e.g. Z0) when not in SSVE/SME mode,
this should return an error.
This does mean that when you're running on an SME capabable machine, but
not in SME mode, and do `register read -a`, lldb will report that 48 SVE
registers were unavailable and 5 SME registers were unavailable. But
that's only when `-a` is used.
The register reading and writing depends on new register flavor support
in thread_get_state/thread_set_state in the kernel, which is not yet in
a release. The test case I wrote is skipped on current OSes. I pilfered
the SME register setup from some of David's existing SME test files;
there were a few Linux specific details in those tests that they weren't
easy to reuse on Darwin.
rdar://121608074
The `llvm-gcc` front-end has been EOL'd at least since 2011 (based on
some `git` archeology). And Clang/LLVM has been removing references to
it ever since.
This patch removes the remaining references to it from LLDB. One benefit
of this is that it will allow us to remove the code checking for
`DW_AT_decl_file_attributes_are_invalid` and
`Supports_DW_AT_APPLE_objc_complete_type`.
With all the recent versions of Clang that I tested, ObjC forward
declarations like
```
@class ForwardObjcClass;
```
don't emit the kind of DWARF that this workaround was put in place for.
Also, zero-sized structures are valid in C (and thus Objective-C), so
this workaround makes things confusing to reason about when mixing the
two languages.
This workaround has been in place for at least a decade, and given that
recent compilers don't produce this anymore, we think it's a good time
to remove it.
For high-frequency multithreaded progress reports, the contention of
taking the progress mutex and the overhead of generating event can
significantly slow down the operation whose progress we are reporting.
This patch adds an (optional) capability to rate-limit them. It's
optional because this approach has one drawback: if the progress
reporting has a pattern where it generates a burst of activity and then
blocks (without reporting anything) for a large amount of time, it can
appear as if less progress was made that it actually was (because we
only reported the first event from the burst and dropped the other
ones).
I've also made a small refactor of the Progress class to better
distinguish between const (don't need protection), atomic (are used on
the hot path) and other (need mutex protection) members.
The main difference is that the llvm class (just a std::vector in
disguise) is not sorted. It turns out this isn't an issue because the
callers either:
- ignore the range list;
- convert it to a different format (which is then sorted);
- or query the minimum value (which is faster than sorting)
The last case is something I want to get rid of in a followup as a part
of removing the assumption that function's entry point is also its
lowest address.
Since the setup of debug registers for AArch64 and LoongArch is similar,
we extracted the shared logic from Class:
`NativeRegisterContextDBReg_arm64`
into a new Class:
`NativeRegisterContextDBReg`.
This will simplify the subsequent implementation of hardware breakpoints
and watchpoints on LoongArch.
Reviewed By: DavidSpickett
Pull Request: https://github.com/llvm/llvm-project/pull/118043
Apologies for the large change, I looked for ways to break this up and
all of the ones I saw added real complexity. This change focuses on the
option's prefixed names and the array of prefixes. These are present in
every option and the dominant source of dynamic relocations for PIE or
PIC users of LLVM and Clang tooling. In some cases, 100s or 1000s of
them for the Clang driver which has a huge number of options.
This PR addresses this by building a string table and a prefixes table
that can be referenced with indices rather than pointers that require
dynamic relocations. This removes almost 7k dynmaic relocations from the
`clang` binary, roughly 8% of the remaining dynmaic relocations outside
of vtables. For busy-boxing use cases where many different option tables
are linked into the same binary, the savings add up a bit more.
The string table is a straightforward mechanism, but the prefixes
required some subtlety. They are encoded in a Pascal-string fashion with
a size followed by a sequence of offsets. This works relatively well for
the small realistic prefixes arrays in use.
Lots of code has to change in order to land this though: both all the
option library code has to be updated to use the string table and
prefixes table, and all the users of the options library have to be
updated to correctly instantiate the objects.
Some follow-up patches in the works to provide an abstraction for this
style of code, and to start using the same technique for some of the
other strings here now that the infrastructure is in place.
Reland https://github.com/llvm/llvm-project/pull/83237
---
(Original comments)
Currently all the specializations of a template (including
instantiation, specialization and partial specializations) will be
loaded at once if we want to instantiate another instance for the
template, or find instantiation for the template, or just want to
complete the redecl chain.
This means basically we need to load every specializations for the
template once the template declaration got loaded. This is bad since
when we load a specialization, we need to load all of its template
arguments. Then we have to deserialize a lot of unnecessary
declarations.
For example,
```
// M.cppm
export module M;
export template <class T>
class A {};
export class ShouldNotBeLoaded {};
export class Temp {
A<ShouldNotBeLoaded> AS;
};
// use.cpp
import M;
A<int> a;
```
We have a specialization ` A<ShouldNotBeLoaded>` in `M.cppm` and we
instantiate the template `A` in `use.cpp`. Then we will deserialize
`ShouldNotBeLoaded` surprisingly when compiling `use.cpp`. And this
patch tries to avoid that.
Given that the templates are heavily used in C++, this is a pain point
for the performance.
This patch adds MultiOnDiskHashTable for specializations in the
ASTReader. Then we will only deserialize the specializations with the
same template arguments. We made that by using ODRHash for the template
arguments as the key of the hash table.
To review this patch, I think `ASTReaderDecl::AddLazySpecializations`
may be a good entry point.
Compared to the python version, this also does type checking and error
handling, so it's slightly longer, however, it's still comfortably
under 500 lines.
Relanding with more explicit type conversions.
This reverts commit f6012a209d.
Revert "[lldb] Add cast to fix compile error on 32-but platforms"
This reverts commit d300337e93.
Revert "[lldb] Improve log message to include missing strings"
This reverts commit 0be3348485.
Revert "[lldb] Add comment"
This reverts commit e2bb47443d.
Revert "[lldb] Implement a formatter bytecode interpreter in C++"
This reverts commit 9a9c1d4a61.
Compared to the python version, this also does type checking and error
handling, so it's slightly longer, however, it's still comfortably
under 500 lines.
Add support for type summaries embedded into the binary.
These embedded summaries will typically be generated by Swift macros,
but can also be generated by any other means.
rdar://115184658
Indexing a single DWARF unit is a fairly small task, which means the
overhead of enqueueing a task for each unit is not negligible (mainly
because introduces a lot of synchronization points for queue management,
memory allocation etc.). This is particularly true if the binary was
built with type units, as these are usually very small.
This essentially brings us back to the state before
https://reviews.llvm.org/D78337, but the new implementation is built on
the llvm ThreadPool, and I've added a small improvement -- we now
construct one "index set" per thread instead of one per unit, which
should lower the memory usage (fewer small allocations) and make the
subsequent merge step faster.
On its own this patch doesn't actually change the performance
characteristics because we still have one choke point -- progress
reporting. I'm leaving that for a separate patch, but I've tried that
simply removing the progress reporting gives us about a 30-60% speed
boost.
LLDB can crash in TypeSystemClang::GetIndexOfChildMemberWithName, at a
point where it pushes an index onto the child_indexes vector, tries to
call itself recursively, then tries to pop the entry from child_indexes.
The problem is that the recursive call can clear child_indexes, so that
this code ends up trying to pop an already empty vector. This change
saves the old vector before the push, then restores the saved vector
rather than trying to pop.
The ManualDWARFIndex class can create a index cache if the LLDB index
cache is enabled. This used to save the index to the same file,
regardless of wether the cache was a full index (no .debug_names) or a
partial index (have .debug_names, but not all .o files were had
.debug_names). So we could end up saving an index cache that was
partial, and then later load that partial index as if it were a full
index if the user set the 'settings set
plugin.symbol-file.dwarf.ignore-file-indexes true'. This would cause us
to ignore the .debug_names section, and if the index cache was enabled,
we could end up loading the partial index as if it were a full DWARF
index.
This patch detects when the ManualDWARFIndex is being used with
.debug_names, and saves out a cache file with a suffix of "-full" or
"-partial" to avoid this issue.
.. in the global namespace
The problem was the interaction of #116989 with an optimization in
GetTypesWithQuery. The optimization was only correct for non-exact
matches, but that didn't matter before this PR due to the "second layer
of defense". After that was removed, the query started returning more
types than it should.
This reverts commit 2526d5b168, reapplying
ba14dac481 after fixing the conflict with
#117532. The change is that Function::GetAddressRanges now recomputes
the returned value instead of returning the member. This means it now
returns a value instead of a reference type.
This is a follow-up/reimplementation of #115730. While working on that
patch, I did not realize that the correct (discontinuous) set of ranges
is already stored in the block representing the whole function. The
catch -- ranges for this block are only set later, when parsing all of
the blocks of the function.
This patch changes that by populating the function block ranges eagerly
-- from within the Function constructor. This also necessitates a
corresponding change in all of the symbol files -- so that they stop
populating the ranges of that block. This allows us to avoid some
unnecessary work (not parsing the function DW_AT_ranges twice) and also
results in some simplification of the parsing code.
The Mach-O load commands have an LC_SYMTAB / struct symtab_command which
represents the offset of the symbol table (nlist records) and string
table for this binary. In a mach-o binary on disk, these are file
offsets. If a mach-o binary is loaded in memory with all segments
consecutive, the `symoff` and `stroff` are the offsets from the TEXT
segment (aka the mach-o header) virtual address to the virtual address
of the start of these tables.
However, if a Mach-O binary is a part of the shared cache, then the
segments will be separated -- they will have different slide values. And
it is possible for the LINKEDIT segment to be greater than 4GB away from
the TEXT segment in the virtual address space, so these 32-bit offsets
cannot express the offset from TEXT segment to these tables.
Create separate uint64_t variables to track the offset to the symbol
table and string table, instead of reusing the 32-bit ones in the
symtab_command structure.
rdar://140432279
It's never set to true. Also, using inheritable FDs in a multithreaded
process pretty much guarantees descriptor leaks. It's better to
explicitly pass a specific FD to a specific subprocess, which we already
mostly can do using the ProcessLaunchInfo FileActions.
It's basically true already (except for a brief time during
construction). This patch makes sure the objects are constructed with a
valid parent and enforces this in the type system, which allows us to
get rid of some nullptr checks.
Allows us to stop waiting for a connection if it doesn't come in a
certain amount of time. Right now, I'm keeping the status quo (infitnite
wait) in the "production" code, but using smaller (finite) values in
tests. (A lot of these tests create "loopback" connections, where a
really short wait is sufficient: on linux at least even a poll (0s wait)
is sufficient if the other end has connect()ed already, but this doesn't
seem to be the case on Windows, so I'm using a 1s wait in these cases).
On #110065 the changes to LinuxSigInfo Struct introduced some variables
that will differ in size on 32b or 64b. I've rectified this by setting
them all to build independent types.
In case of an error GetBlock would return a reference to a Block without
adding it to a parent. This doesn't seem like a good idea, and none of
the other plugins do that.
This patch fixes that by propagating errors (well, null pointers...) up
the stack.
I don't know of any specific problem that this solves, but given that
this occurs only when something goes very wrong (e.g. a corrupted PDB
file), it's quite possible noone has run into this situation, so we
can't say the code is correct either. It also gets in the way of a
refactor I'm contemplating.
In #108907, the index classes started filtering the DIEs according to
the full type query (instead of just the base name). This means that the
checks in SymbolFileDWARF are now redundant.
I've also moved the non-redundant checks so that now all checking is
done in the DWARFIndex class and the caller can expect to get the final
filtered list of types.
ELF core debugging fix#117070 broke TestLoadUnload.py tests due to
GetModuleSpec call, ProcessGDBRemote fetches modules from remote. Revise
the original PR, renamed FindBuildId to FindModuleUUID.
This patch fixes:
lldb/source/Plugins/Process/elf-core/ThreadElfCore.cpp:53:32: error:
field 'm_thread_reg_ctx_sp' will be initialized after field
'm_thread_name' [-Werror,-Wreorder-ctor]
Prior to this patch, the function returned an exit status, sometimes a
ValueObject with an error and a Status object. This patch removes the
Status object and ensures the error is consistently returned as the
error of the ValueObject.
This fixes a functionality gap with GDB, where GDB will properly decode
the stop reason and give the address for SIGSEGV. I also added
descriptions to all stop reasons, following the same code path that the
Native Linux Thread uses.
A previous patch added the ability to load UUID from ELF headers using
the program header and finding PT_NOTE entries. The fix would attempt to
read the data for the PT_NOTE from memory, but it didn't slide the
address so it ended up only working for the main executable if it wasn't
moved in memory. This patch slides the address and adds logging.
All processes map the ELF header + program headers + some program header
contents into memory. The program header for the `PT_NOTE` entries are
mapped, but the p_vaddr doesn't get relocated and is relative to the
load address of the ELF header. So we take a "p_vaddr" (file address)
and convert it into a load address in the process so we can load the
correct bytes that contain the `PT_NOTE` contents.
I backed this out due to a problem on one of the bots that myself and
others have problems reproducing locally. I'd like to try to land it
again, at least to gain more information.
Summary:
This improves the performance of ObjectFileMacho::ParseSymtab by
removing eager and expensive work in favor of doing it later in a
less-expensive fashion.
Experiment:
My goal was to understand LLDB's startup time.
First, I produced a Debug build of LLDB (no dSYM) and a
Release+NoAsserts build of LLDB. The Release build debugged the Debug
build as it debugged a small C++ program. I found that
ObjectFileMachO::ParseSymtab accounted for somewhere between 1.2 and 1.3
seconds consistently. After applying this change, I consistently
measured a reduction of approximately 100ms, putting the time closer to
1.1s and 1.2s on average.
Background:
ObjectFileMachO::ParseSymtab will incrementally create symbols by
parsing nlist entries from the symtab section of a MachO binary. As it
does this, it eagerly tries to determine the size of symbols (e.g. how
long a function is) using LC_FUNCTION_STARTS data (or eh_frame if
LC_FUNCTION_STARTS is unavailable). Concretely, this is done by
performing a binary search on the function starts array and calculating
the distance to the next function or the end of the section (whichever
is smaller).
However, this work is unnecessary for 2 reasons:
1. If you have debug symbol entries (i.e. STABs), the size of a function
is usually stored right after the function's entry. Performing this work
right before parsing the next entry is unnecessary work.
2. Calculating symbol sizes for symbols of size 0 is already performed
in `Symtab::InitAddressIndexes` after all the symbols are added to the
Symtab. It also does this more efficiently by walking over a list of
symbols sorted by address, so the work to calculate the size per symbol
is constant instead of O(log n).
This is motivated by exposing some Swift language-specific flags through
the API, in the example here it is used to communicate the Objective-C
runtime version. This could also be a meaningful extension point to get
information about "embedded: languages, such as extracting the C++
version in an Objective-C++ frame or something along those lines.
Resubmissions of https://github.com/llvm/llvm-project/pull/112596 with
buildbot fixes.
Allow LLDB to parse the dynamic symbol table from an ELF file or memory
image in an ELF file that has no section headers. This patch uses the
ability to parse the PT_DYNAMIC segment and find the DT_SYMTAB,
DT_SYMENT, DT_HASH or DT_GNU_HASH to find and parse the dynamic symbol
table if the section headers are not present. It also adds a helper
function to read data from a .dynamic key/value pair entry correctly
from the file or from memory.
