D104422 added the interface for TraceCursor, which is the main way to traverse instructions in a trace. This diff implements the corresponding cursor class for Intel PT and deletes the now obsolete code.
Besides that, the logic for the "thread trace dump instructions" was adapted to use this cursor (pretty much I ended up moving code from Trace.cpp to TraceCursor.cpp). The command by default traverses the instructions backwards, and if the user passes --forwards, then it's not forwards. More information about that is in the Options.td file.
Regarding the Intel PT cursor. All Intel PT cursors for the same thread share the same DecodedThread instance. I'm not yet implementing lazy decoding because we don't need it. That'll be for later. For the time being, the entire thread trace is decoded when the first cursor for that thread is requested.
Differential Revision: https://reviews.llvm.org/D105531
We've seen reports of crashes (none we've been able to reproduce
locally) that look like they are caused by concurrent access to a
thread plan stack. It looks like there are error paths when an
interrupt request to debugserver times out that cause this problem.
The thread plan stack access is never in a hot loop, and there
aren't enough of them for the extra data member to matter, so
there's really no good reason not to protect the access.
Adding the mutex revealed a couple of places where we were
using "auto" in an iteration when we should have been using
"auto &" - we didn't intend to copy the stack - and I fixed
those as well.
Except for preventing crashes this should be NFC.
Differential Revision: https\://reviews.llvm.org/D106122
PackTags is used by to compress tags to go in the QMemTags packet
and be passed to ptrace when writing memory tags.
The behaviour of RepeatTagsForRange matches that described for QMemTags
in the GDB documentation:
https://sourceware.org/gdb/current/onlinedocs/gdb/General-Query-Packets.html#General-Query-Packets
In addition, unpacking tags with number of tags 0 now means
do not check that number of tags matches the range.
This will be used by lldb-server to unpack tags before repeating
them to fill the requested range.
Reviewed By: omjavaid
Differential Revision: https://reviews.llvm.org/D105179
Previously GetMemoryTagManager checked many things in one:
* architecture supports memory tagging
* process supports memory tagging
* memory range isn't inverted
* memory range is all tagged
Since writing follow up patches for tag writing (in review
at the moment) it has become clear that this gets unwieldy
once we add the features needed for that.
It also implies that the memory tag manager is tied to the
range you used to request it with but it is not. It's a per
process object.
Instead:
* GetMemoryTagManager just checks architecture and process.
* Then the MemoryTagManager can later be asked to check a
memory range.
This is better because:
* We don't imply that range and manager are tied together.
* A slightly diferent range calculation for tag writing
doesn't add more code to Process.
* Range checking code can now be unit tested.
Reviewed By: omjavaid
Differential Revision: https://reviews.llvm.org/D105630
This change adds AllocateMemory and DeallocateMemory methods to the SBProcess
API, so that clients can allocate and deallocate memory blocks within the
process being debugged (for storing JIT-compiled code or other uses).
(I am developing a debugger + REPL using the API; it will need to store
JIT-compiled code within the target.)
Reviewed By: clayborg, jingham
Differential Revision: https://reviews.llvm.org/D105389
This reverts commit 82a38837150099288a1262391ef43e1fd69ffde4.
The original version had a copy-paste error: using the Interrupt timeout
for the ResumeSynchronous wait, which is clearly wrong. This error would
have been evident with real use, but the interrupt is long enough that it
only caused one testsuite failure (in the Swift fork).
Anyway, I found that mistake and fixed it and checked all the other places
where I had to plumb through a timeout, and added a test with a short
interrupt timeout stepping over a function that takes 3x the interrupt timeout
to complete, so that should detect a similar mistake in the future.
Add the ability to silence command script import. The motivation for
this change is being able to add command script import -s
lldb.macosx.crashlog to your ~/.lldbinit without it printing the
following message at the beginning of every debug session.
"malloc_info", "ptr_refs", "cstr_refs", "find_variable", and
"objc_refs" commands have been installed, use the "--help" options on
these commands for detailed help.
In addition to forwarding the silent option to LoadScriptingModule, this
also changes ScriptInterpreterPythonImpl::ExecuteOneLineWithReturn and
ScriptInterpreterPythonImpl::ExecuteMultipleLines to honor the enable IO
option in ExecuteScriptOptions, which until now was ignored.
Note that IO is only enabled (or disabled) at the start of a session,
and for this particular use case, that's done when taking the Python
lock in LoadScriptingModule, which means that the changes to these two
functions are not strictly necessary, but (IMO) desirable nonetheless.
Differential revision: https://reviews.llvm.org/D105327
This reverts commit c8164d0276b97679e80db01adc860271ab4a5d11 and
43f6dad2344247976d5777f56a1fc29e39c6c717 because it breaks
TestDyldTrieSymbols.py on GreenDragon.
This fix was created after profiling the target creation of a large C/C++/ObjC application that contained almost 4,000,000 redacted symbol names. The symbol table parsing code was creating names for each of these synthetic symbols and adding them to the name indexes. The code was also adding the object file basename to the end of the symbol name which doesn't allow symbols from different shared libraries to share the names in the constant string pool.
Prior to this fix this was creating 180MB of "___lldb_unnamed_symbol" symbol names and was taking a long time to generate each name, add them to the string pool and then add each of these names to the name index.
This patch fixes the issue by:
- not adding a name to synthetic symbols at creation time, and allows name to be dynamically generated when accessed
- doesn't add synthetic symbol names to the name indexes, but catches this special case as name lookup time. Users won't typically set breakpoints or lookup these synthetic names, but support was added to do the lookup in case it does happen
- removes the object file baseanme from the generated names to allow the names to be shared in the constant string pool
Prior to this fix the startup times for a large application was:
12.5 seconds (cold file caches)
8.5 seconds (warm file caches)
After this fix:
9.7 seconds (cold file caches)
5.7 seconds (warm file caches)
The names of the symbols are auto generated by appending the symbol's UserID to the end of the "___lldb_unnamed_symbol" string and is only done when the name is requested from a synthetic symbol if it has no name.
Differential Revision: https://reviews.llvm.org/D105160
This is an NFC modernization refactoring that replaces the combination
of a bool return + reference argument, with an Optional return value.
Differential Revision: https://reviews.llvm.org/D104405
Reverts commits:
"Fix failing tests after https://reviews.llvm.org/D104488."
"Fix buildbot failure after https://reviews.llvm.org/D104488."
"Create synthetic symbol names on demand to improve memory consumption and startup times."
This series of commits broke the windows lldb bot and then failed to fix all of the failing tests.
This patch introduces a new interpreter setting
`interpreter.save-session-directory` so the user can specify a directory
where the session transcripts will be saved.
If not set, the session transcript are saved on a temporary file.
rdar://72902842
Differential Revision: https://reviews.llvm.org/D105030
Signed-off-by: Med Ismail Bennani <medismail.bennani@gmail.com>
This fix was created after profiling the target creation of a large C/C++/ObjC application that contained almost 4,000,000 redacted symbol names. The symbol table parsing code was creating names for each of these synthetic symbols and adding them to the name indexes. The code was also adding the object file basename to the end of the symbol name which doesn't allow symbols from different shared libraries to share the names in the constant string pool.
Prior to this fix this was creating 180MB of "___lldb_unnamed_symbol" symbol names and was taking a long time to generate each name, add them to the string pool and then add each of these names to the name index.
This patch fixes the issue by:
- not adding a name to synthetic symbols at creation time, and allows name to be dynamically generated when accessed
- doesn't add synthetic symbol names to the name indexes, but catches this special case as name lookup time. Users won't typically set breakpoints or lookup these synthetic names, but support was added to do the lookup in case it does happen
- removes the object file baseanme from the generated names to allow the names to be shared in the constant string pool
Prior to this fix the startup times for a large application was:
12.5 seconds (cold file caches)
8.5 seconds (warm file caches)
After this fix:
9.7 seconds (cold file caches)
5.7 seconds (warm file caches)
The names of the symbols are auto generated by appending the symbol's UserID to the end of the "___lldb_unnamed_symbol" string and is only done when the name is requested from a synthetic symbol if it has no name.
Differential Revision: https://reviews.llvm.org/D104488
This is an NFC modernization refactoring that replaces the combination
of a bool return + reference argument, with an Optional return value.
Differential Revision: https://reviews.llvm.org/D104404
This adds GDB client support for the qMemTags packet
which reads memory tags. Following the design
which was recently committed to GDB.
https://sourceware.org/gdb/current/onlinedocs/gdb/General-Query-Packets.html#General-Query-Packets
(look for qMemTags)
lldb commands will use the new Process methods
GetMemoryTagManager and ReadMemoryTags.
The former takes a range and checks that:
* The current process architecture has an architecture plugin
* That plugin provides a MemoryTagManager
* That the range of memory requested lies in a tagged range
(it will expand it to granules for you)
If all that was true you get a MemoryTagManager you
can give to ReadMemoryTags.
This two step process is done to allow commands to get the
tag manager without having to read tags as well. For example
you might just want to remove a logical tag, or error early
if a range with tagged addresses is inverted.
Note that getting a MemoryTagManager doesn't mean that the process
or a specific memory range is tagged. Those are seperate checks.
Having a tag manager just means this architecture *could* have
a tagging feature enabled.
An architecture plugin has been added for AArch64 which
will return a MemoryTagManagerAArch64MTE, which was added in a
previous patch.
Reviewed By: omjavaid
Differential Revision: https://reviews.llvm.org/D95602
This adds memory tag reading using the new "qMemTags"
packet and ptrace on AArch64 Linux.
This new packet is following the one used by GDB.
(https://sourceware.org/gdb/current/onlinedocs/gdb/General-Query-Packets.html)
On AArch64 Linux we use ptrace's PEEKMTETAGS to read
tags and we assume that lldb has already checked that the
memory region actually has tagging enabled.
We do not assume that lldb has expanded the requested range
to granules and expand it again to be sure.
(although lldb will be sending aligned ranges because it happens
to need them client side anyway)
Also we don't assume untagged addresses. So for AArch64 we'll
remove the top byte before using them. (the top byte includes
MTE and other non address data)
To do the ptrace read NativeProcessLinux will ask the native
register context for a memory tag manager based on the
type in the packet. This also gives you the ptrace numbers you need.
(it's called a register context but it also has non register data,
so it saves adding another per platform sub class)
The only supported platform for this is AArch64 Linux and the only
supported tag type is MTE allocation tags. Anything else will
error.
Ptrace can return a partial result but for lldb-server we will
be treating that as an error. To succeed we need to get all the tags
we expect.
(Note that the protocol leaves room for logical tags to be
read via qMemTags but this is not going to be implemented for lldb
at this time.)
Reviewed By: omjavaid
Differential Revision: https://reviews.llvm.org/D95601
This feature "memory-tagging+" indicates that lldb-server
supports memory tagging packets. (added in a later patch)
We check HWCAP2_MTE to decide whether to enable this
feature for Linux.
Reviewed By: omjavaid
Differential Revision: https://reviews.llvm.org/D97282
This adds the MemoryTagManager class and a specialisation
of that class for AArch64 MTE tags. It provides a generic
interface for various tagging operations.
Adding/removing tags, diffing tagged pointers, etc.
Later patches will use this manager to handle memory tags
in generic code in both lldb and lldb-server.
Since it will be used in both, the base class header is in
lldb/Target.
(MemoryRegionInfo is another example of this pattern)
Reviewed By: omjavaid
Differential Revision: https://reviews.llvm.org/D97281
As a follow up of D103588, I'm reinitiating the discussion with a new proposal for traversing instructions in a trace which uses the feedback gotten in that diff.
See the embedded documentation in TraceCursor for more information. The idea is to offer an OOP way to traverse instructions exposing a minimal interface that makes no assumptions on:
- the number of instructions in the trace (i.e. having indices for instructions might be impractical for gigantic intel-pt traces, as it would require to decode the entire trace). This renders the use of indices to point to instructions impractical. Traces are big and expensive, and the consumer should try to do look linear lookups (forwards and/or backwards) and avoid random accesses (the API could be extended though, but for now I want to dicard that funcionality and leave the API extensible if needed).
- the way the instructions are represented internally by each Trace plug-in. They could be mmap'ed from a file, exist in plain vector or generated on the fly as the user requests the data.
- the actual data structure used internally for each plug-in. Ideas like having a struct TraceInstruction have been discarded because that would make the plug-in follow a certain data type, which might be costly. Instead, the user can ask the cursor for each independent property of the instruction it's pointing at.
The way to get a cursor is to ask Trace.h for the end or being cursor or a thread's trace.
There are some benefits of this approach:
- there's little cost to create a cursor, and this allows for lazily decoding a trace as the user requests data.
- each trace plug-in could decide how to cache the instructions it generates. For example, if a trace is small, it might decide to keep everything in memory, or if the trace is massive, it might decide to keep around the last thousands of instructions to speed up local searches.
- a cursor can outlive a stop point, which makes trace comparison for live processes feasible. An application of this is to compare profiling data of two runs of the same function, which should be doable with intel pt.
Differential Revision: https://reviews.llvm.org/D104422
We can extend/modify `GetMethodNameVariants` to suit our purposes here.
What symtab is looking for is alternate names we may want to use to
search for a specific symbol, and asking for variants of a name makes
the most sense here.
Differential Revision: https://reviews.llvm.org/D104067
Replacing existing uses with AppendError.
SetError is also part of the SBI API. This remains
but instead of calling the underlying SetError it
will call AppendError.
Reviewed By: teemperor
Differential Revision: https://reviews.llvm.org/D104768
Rust's v0 name mangling scheme [1] is easy to disambiguate from other
name mangling schemes because symbols always start with `_R`. The llvm
Demangle library supports demangling the Rust v0 scheme. Use it to
demangle Rust symbols.
Added unit tests that check simple symbols. Ran LLDB built with this
patch to debug some Rust programs compiled with the v0 name mangling
scheme. Confirmed symbol names were demangled as expected.
Note: enabling the new name mangling scheme requires a nightly
toolchain:
```
$ cat main.rs
fn main() {
println!("Hello world!");
}
$ $(rustup which --toolchain nightly rustc) -Z symbol-mangling-version=v0 main.rs -g
$ /home/asm/hacking/llvm/build/bin/lldb ./main --one-line 'b main.rs:2'
(lldb) target create "./main"
Current executable set to '/home/asm/hacking/llvm/rust/main' (x86_64).
(lldb) b main.rs:2
Breakpoint 1: where = main`main::main + 4 at main.rs:2:5, address = 0x00000000000076a4
(lldb) r
Process 948449 launched: '/home/asm/hacking/llvm/rust/main' (x86_64)
warning: (x86_64) /lib64/libgcc_s.so.1 No LZMA support found for reading .gnu_debugdata section
Process 948449 stopped
* thread #1, name = 'main', stop reason = breakpoint 1.1
frame #0: 0x000055555555b6a4 main`main::main at main.rs:2:5
1 fn main() {
-> 2 println!("Hello world!");
3 }
(lldb) bt
error: need to add support for DW_TAG_base_type '()' encoded with DW_ATE = 0x7, bit_size = 0
* thread #1, name = 'main', stop reason = breakpoint 1.1
* frame #0: 0x000055555555b6a4 main`main::main at main.rs:2:5
frame #1: 0x000055555555b78b main`<fn() as core::ops::function::FnOnce<()>>::call_once((null)=(main`main::main at main.rs:1), (null)=<unavailable>) at function.rs:227:5
frame #2: 0x000055555555b66e main`std::sys_common::backtrace::__rust_begin_short_backtrace::<fn(), ()>(f=(main`main::main at main.rs:1)) at backtrace.rs:125:18
frame #3: 0x000055555555b851 main`std::rt::lang_start::<()>::{closure#0} at rt.rs:49:18
frame #4: 0x000055555556c9f9 main`std::rt::lang_start_internal::hc51399759a90501a [inlined] core::ops::function::impls::_$LT$impl$u20$core..ops..function..FnOnce$LT$A$GT$$u20$for$u20$$RF$F$GT$::call_once::h04259e4a34d07c2f at function.rs:259:13
frame #5: 0x000055555556c9f2 main`std::rt::lang_start_internal::hc51399759a90501a [inlined] std::panicking::try::do_call::hb8da45704d5cfbbf at panicking.rs:401:40
frame #6: 0x000055555556c9f2 main`std::rt::lang_start_internal::hc51399759a90501a [inlined] std::panicking::try::h4beadc19a78fec52 at panicking.rs:365:19
frame #7: 0x000055555556c9f2 main`std::rt::lang_start_internal::hc51399759a90501a [inlined] std::panic::catch_unwind::hc58016cd36ba81a4 at panic.rs:433:14
frame #8: 0x000055555556c9f2 main`std::rt::lang_start_internal::hc51399759a90501a at rt.rs:34:21
frame #9: 0x000055555555b830 main`std::rt::lang_start::<()>(main=(main`main::main at main.rs:1), argc=1, argv=0x00007fffffffcb18) at rt.rs:48:5
frame #10: 0x000055555555b6fc main`main + 28
frame #11: 0x00007ffff73f2493 libc.so.6`__libc_start_main + 243
frame #12: 0x000055555555b59e main`_start + 46
(lldb)
```
[1]: https://github.com/rust-lang/rust/issues/60705
Reviewed By: clayborg, teemperor
Differential Revision: https://reviews.llvm.org/D104054
HostInfoBase has a deleted dtor/ctor so there is no need to do the same for
all the classes inheriting from it.
Reviewed By: DavidSpickett, JDevlieghere
Differential Revision: https://reviews.llvm.org/D104221
Add a new feature to process save-core on Darwin systems -- for
lldb to create a user process corefile with only the dirty (modified
memory) pages included. All of the binaries that were used in the
corefile are assumed to still exist on the system for the duration
of the use of the corefile. A new --style option to process save-core
is added, so a full corefile can be requested if portability across
systems, or across time, is needed for this corefile.
debugserver can now identify the dirty pages in a memory region
when queried with qMemoryRegionInfo, and the size of vm pages is
given in qHostInfo.
Create a new "all image infos" LC_NOTE for Mach-O which allows us
to describe all of the binaries that were loaded in the process --
load address, UUID, file path, segment load addresses, and optionally
whether code from the binary was executing on any thread. The old
"read dyld_all_image_infos and then the in-memory Mach-O load
commands to get segment load addresses" no longer works when we
only have dirty memory.
rdar://69670807
Differential Revision: https://reviews.llvm.org/D88387
This adds a basic SB API for creating and stopping traces.
Note: This doesn't add any APIs for inspecting individual instructions. That'd be a more complicated change and it might be better to enhande the dump functionality to output the data in binary format. I'll leave that for a later diff.
This also enhances the existing tests so that they test the same flow using both the command interface and the SB API.
I also did some cleanup of legacy code.
Differential Revision: https://reviews.llvm.org/D103500
This is an NFC cleanup.
Many of the API's that returned BreakpointOptions always returned valid ones.
Internally the BreakpointLocations usually have null BreakpointOptions, since they
use their owner's options until an option is set specifically on the location.
So the original code used pointers & unique_ptr everywhere for consistency.
But that made the code hard to reason about from the outside.
This patch changes the code so that everywhere an API is guaranteed to
return a non-null BreakpointOption, it returns it as a reference to make
that clear.
It also changes the Breakpoint to hold a BreakpointOption
member where it previously had a UP. Since we were always filling the UP
in the Breakpoint constructor, having the UP wasn't helping anything.
Differential Revision: https://reviews.llvm.org/D104162
One nice feature of the os_signpost API is that format string
substitutions happen in the consumer, not the logging
application. LLVM's current Signpost class doesn't take advantage of
this though and instead always uses a static "Begin/End %s" format
string.
This patch uses variadic macros to allow the API to be used as
intended. Unfortunately, the primary use-case I had in mind (the
LLDB_SCOPED_TIMER() macro) does not get much better from this, because
__PRETTY_FUNCTION__ is *not* a macro, but a static string, so
signposts created by LLDB_SCOPED_TIMER() still use a static "%s"
format string. At least LLDB_SCOPED_TIMERF() works as intended.
This reapplies the previously reverted patch with additional include
order fixes for non-modular builds of LLDB.
Differential Revision: https://reviews.llvm.org/D103575
One nice feature of the os_signpost API is that format string
substitutions happen in the consumer, not the logging
application. LLVM's current Signpost class doesn't take advantage of
this though and instead always uses a static "Begin/End %s" format
string.
This patch uses variadic macros to allow the API to be used as
intended. Unfortunately, the primary use-case I had in mind (the
LLDB_SCOPED_TIMER() macro) does not get much better from this, because
__PRETTY_FUNCTION__ is *not* a macro, but a static string, so
signposts created by LLDB_SCOPED_TIMER() still use a static "%s"
format string. At least LLDB_SCOPED_TIMERF() works as intended.
This reapplies the previsously reverted patch with additional MachO.h
macro #undefs.
Differential Revision: https://reviews.llvm.org/D103575
One nice feature of the os_signpost API is that format string
substitutions happen in the consumer, not the logging
application. LLVM's current Signpost class doesn't take advantage of
this though and instead always uses a static "Begin/End %s" format
string.
This patch uses variadic macros to allow the API to be used as
intended. Unfortunately, the primary use-case I had in mind (the
LLDB_SCOPED_TIMER() macro) does not get much better from this, because
__PRETTY_FUNCTION__ is *not* a macro, but a static string, so
signposts created by LLDB_SCOPED_TIMER() still use a static "%s"
format string. At least LLDB_SCOPED_TIMERF() works as intended.
This reapplies the previsously reverted patch with support for
platforms where signposts are unavailable.
Differential Revision: https://reviews.llvm.org/D103575
One nice feature of the os_signpost API is that format string
substitutions happen in the consumer, not the logging
application. LLVM's current Signpost class doesn't take advantage of
this though and instead always uses a static "Begin/End %s" format
string.
This patch uses variadic macros to allow the API to be used as
intended. Unfortunately, the primary use-case I had in mind (the
LLDB_SCOPED_TIMER() macro) does not get much better from this, because
__PRETTY_FUNCTION__ is *not* a macro, but a static string, so
signposts created by LLDB_SCOPED_TIMER() still use a static "%s"
format string. At least LLDB_SCOPED_TIMERF() works as intended.
Differential Revision: https://reviews.llvm.org/D103575
Test leaks if we run
tools/lldb/unittests/Host/HostTests without --gtest_filter
Reviewed By: teemperor
Differential Revision: https://reviews.llvm.org/D104091
This converts a default constructor's member initializers into C++11
default member initializers. This patch was automatically generated with
clang-tidy and the modernize-use-default-member-init check.
$ run-clang-tidy.py -header-filter='lldb' -checks='-*,modernize-use-default-member-init' -fix
This is a mass-refactoring patch and this commit will be added to
.git-blame-ignore-revs.
Differential revision: https://reviews.llvm.org/D103483
When executing a script command in HandleCommand(s) we currently print
its output twice
You can see this issue in action when adding a breakpoint command:
(lldb) b main
Breakpoint 1: where = main.out`main + 13 at main.cpp:2:3, address = 0x0000000100003fad
(lldb) break command add 1 -o "script print(\"Hey!\")"
(lldb) r
Process 76041 launched: '/tmp/main.out' (x86_64)
Hey!
(lldb) script print("Hey!")
Hey!
Process 76041 stopped
The issue is caused by HandleCommands using a temporary
CommandReturnObject and one of the commands (`script` in this case)
setting an immediate output stream. This causes the result to be printed
twice: once directly to the immediate output stream and once when
printing the result of HandleCommands.
This patch fixes the issue by introducing a new option to suppress
immediate output for temporary CommandReturnObjects.
Differential revision: https://reviews.llvm.org/D103349
The various maps in Symtab lead to some repetative code. This should
improve the situation somewhat.
Differential Revision: https://reviews.llvm.org/D103652
This is another step towards implementing the equivalent of
`platform process list` and related functionality.
`uint32_t` is used for the argument count and index despite the
underlying value being `size_t` to be consistent with other
index-based access to arguments.
Differential Revision: https://reviews.llvm.org/D103675
This is present when doing a `platform process list` and is
tracked by the underlying code. To do something like the
process list via the SB API in the future, this must be
exposed.
Differential Revision: https://reviews.llvm.org/D103375
Previously ignore counts were checked when we stopped to do the sync callback in Breakpoint::ShouldStop. That meant we would do all the ignore count work even when
there is also a condition says the breakpoint should not stop.
That's wrong, lldb treats breakpoint hits that fail the thread or condition checks as "not having hit the breakpoint". So the ignore count check should happen after
the condition and thread checks in StopInfoBreakpoint::PerformAction.
The one side-effect of doing this is that if you have a breakpoint with a synchronous callback, it will run the synchronous callback before checking the ignore count.
That is probably a good thing, since this was already true of the condition and thread checks, so this removes an odd asymmetry. And breakpoints with sync callbacks
are all internal lldb breakpoints and there's not a really good reason why you would want one of these to use an ignore count (but not a condition or thread check...)
Differential Revision https://reviews.llvm.org/D103217
