Age | Commit message (Collapse) | Author | Files | Lines |
|
* Remove dead code in BitArray
* Replace size_t with uint32_t in BitArray
|
|
* Fix ARM/ARM64 hijacking in tail calls
This change fixes an issue that can happen when a function that has tail
calls is hijacked. There are two potential issues:
1. When a function that tail calls another one is hijacked, the LR may be
stored at a different location in the stack frame of the tail call
target.
So just by performing tail call, the hijacked location becomes invalid and
unhijacking would corrupt stack by writing to that location.
2. There is a small window after the caller pops LR from the stack in its
epilog and before the tail called function pushes LR in its prolog when
the hijacked return address would not be not on the stack and so we would
not be able to unhijack.
The fix is to prevent hijacking of functions that contain tail calls.
* Enable the tailcall hijacking test for ARM64
The test JIT/Methodical/tailcall_v4/hijacking should be passing now on
ARM64.
|
|
MEASURE_INFO switch now depends on _DEBUG switch. Ensure NumUntracked,
UntrackedSlotSize, NumUntrackedSize and EncPreservedSlots fields handled correctly in
GcInfoSize::operator+=() and GCInfoSize::Log().
fix #6008
|
|
|
|
Ref #4379
This change implements GcInfo version 2 for all platforms that use the
GcInfo library (all architectures other than X86).
Changes are:
1) Defines ReturnKind enumeration for all platforms
2) Change the GcInfo encoder library to encode the ReturnKind and ReversePInvokeFrame slot
3) Change the CM's GcInfo decoder to encode the ReturnKind and ReversePInvokeFrame slot for GCINFO_VERSION 2
4) Some corrections to GCINFO_MEASUREments
5) Changes to RYU Jit to provide the correct information to the encoder
6) Changes to the VM to use the ReturnKind information while hijacking a thread
- If ReturnKind is available from GcInfo, new hijack routines are used
- Otherwise, fall back to old method (for compatibility)
7) Rework and simplify the thread hijack routines by scanning HijackFrames directly for gcroots
8) Supporting code to implement the above features.
Returning Structs in multiple registers
Hijacking for StructInRegs is currently only implemented for
Unix SystemV ABI Multi-reg struct returns. However, the hijack-workers that use
ReturnKind are ready to handle other platforms (ex: ARM/ARM64 Windows)
once the corresponding HijackTripThread() assembly routines are defined.
The New feature flag: FEATURE_MULTIREG_RETURN is set for platforms where a struct value
can be returned in multiple registers [ex: Windows/Unix ARM/ARM64, Unix-AMD64]
FEATURE_UNIX_AMD64_STRUCT_PASSING is a specific kind of FEATURE_MULTIREG_RETURN
specified by SystemV ABI for AMD64
Compatibility with other JITs
- All new GCInfo generated by RYU Jit is in GcInfo version 2
- All Ngen images must be regenerated with the new GcInfo version.
- Ready-to-run images with old GcInfo will continue to work.
- Jit64/X64 uses the GcInfo library, so it generates GcInfo version 2.
However, it doesn't (yet) provide the data to encode the correct ReturnKind
Similar is the case for ARM32 code running on JIT32, and any other JITs
that may be using GcInfo library but not yet modified to use the new API.
So, compatibility is achived using RT_Unset flag.
When ReturnKind is RT_Unset, it means that the JIT did not set
the ReturnKind in the GCInfo, and therefore the VM cannot rely on it,
and must use other mechanisms (similar to GcInfo ver 1) to determine
the Return type's GC information.
Implement GC root scanning for Hijack-frames
This change implements GCScanRoots() method for Hijacke-frames
based on the ReturnKind information available from the GcInfo.
If the exact ReturnKind is not available in the GcInfo, the
thread-suspension logic will compute the ReturnKind based on
the method-signature.
As a result of this change, several hijack-helpers in the VM
are cleaned up. There's only one implementation of HijackWorker()
to handle all returnKinds.
This change also simplifies the thread-hijack logic by using a
single assembly helper OnHijackTripThread() in most cases.
The only other helper used is for X86 floating point return values
for save/restoring the top of the FP stack.
ARM64
Only GcIndfo v2 is reliably supported for ARM64 platform.
The changes to thread-hijack mechanism fixes #6494 for ARM64.
No measurable change in JIT throughput, performance or native-image
size from this change.
|
|
This change removes the redundant set of GcInfo encoder/decoder
in the CoreCLR tree called Debug encoder/decoders.
These components are expected to be are well-tested versions
for verification, but are not actively used.
They cause additional overhead wrt maintaining two versions of
GcInfo encoder/decoder as GcInfo format changes.
|
|
This required two other minor changes:
- A prototype for `ThrowOutOfMemory` is now defined in gcinfoencoder.h
- `BitArray` now defines an overload of `operator new` that accepts an
`IAllocator`.
|
|
- Encapsulate link management into `MemoryBlockDescList`.
- Replace `MemoryBlockDesc` and its separately-allocated buffer
with a new type, `MemoryBlock`, that stores the link to the
next memory block with the buffer itself.
|
|
Instead of using `SList`, manage the links manually. There are only
a few list manipulations in the encoders, so this is not too onerous.
|
|
- Move the code into the GC info encoder, as that is its
only consumer.
- Remove contracts
- Delete dead code
- Reformat and refactor to current coding standards
|
|
This removes another utilcode dependency from the GC info encoders.
|
|
These behaviors override the default out-of-memory handling
s.t. it is appropriate for the JIT.
|
|
There are a number of macros, method bodies, and type definitions that
do not need to be in gcinfoencoder.h. Instead, these definitions have
been moved into gcinfoencoder.cpp.
|
|
|
|
Fix a stale comment about stack-slot-base
in the GcInfo Encoder interface.
|
|
|
|
The STANDALONE_BUILD switch can be used to build the GcInfoEncoder library
independently by clients outside the CoreClr tree.
The GcInfo library uses some custom data-structures (ex: ArrayList)
and includes some utility libraries (ex: UtilCode) which pull in several other
headers with considerable unrelated content. Rather than porting all the
utility code to suite other clients, the STANDALONE_BUILD switch can be used
to include only the minimal set of headers specific to GcInfo encodings.
Clients of STANDALONE_BUILD will likely use standard library
implementations of data-structures like ArrayList, HashMap etc., in place
of the custom implementation currently used by GcInfoEncoder.
Rather than spew the GcInfoEnoder code with
#ifdef STANDALONE_BUILD ... #else .. #endif blocks, we include a special
header GcInfoUtil.h in STANDALONE_BUILD mode. GcInfoUtil.h is expected to
supply the interface/implementation for the data-structures and utilities
used by GcInfoEncoder. This header should be provided by the clients doing
the standalone build in their source tree.
|
|
This change fixes the following warnings:
1) Assignment in a condition should be wrapped in ()
2) Priority of && / || should be indicated by parentheses.
3) Unknown #pragma optimize ifdefed out for non MSVC
4) Unused functions deleted or put under #ifdef
5) Extra tokens warning disabling moved to the CMakeLists.txt in the src/inc
6) Self assignment of a member or local variable
7) Assigning ~0 to a bitfield member that was just 8 bit wide
It also fixes a bug in the STRESS_LOGxx macro invocation in exceptionhandling.cpp and
stackwalk.cpp related to recent adding the DBG_ADDR macro usage. This macro expanded
a single parameter into two expressions to extract high / low 32 bits separately.
But the STRESS_LOGxx parameters are passed to the StressLog::LogMsg method as follows:
(void*)(size_t)(data1)
That means that the expanded pair x, y would be inserted as data 1 and that leads
to ignoring the x due to the comma operator.
|
|
[tfs-changeset: 1407945]
|