GitLab

Unlike Thumb, Morello uses the LSB as part of the
relocation computation, and we shouldn't set the LSB
on the symbol difference result.

Function size computation now needs to use a local label
to avoid subtracting 1 due to the LSB on the STT_FUNC
symbol.

This ensures Morello gets the various bug fixes and improvements made to
the normal CHERI getTargetSize since it made its own copy. As part of
this, fix the hacky handling of symbol sizes added for Morello as
follows.

getSize only respected the isSectionStartSymbol field for CheriABI, but
hybrid Morello binaries can have capabilities too, so add a new boolean
parameter for CheriCapRelocsSection to use in order to force it to be
honoured.

Linker scripts were also not setting isSectionStartSymbol for things
that looked like section start symbols, so getSize would never use the
section's bounds for them.

Because of these two issues, getMorelloTargetSize overrode a value of
0 for symbols that had a name that looked like a section start symbol
rather than taking into account the additional restrictions from Writer
on whether a symbol is a section start symbol. By ensuring the symbol
information is right in the first place this hack is no longer needed.

Note: isSectionStartSymbol is both a field in Defined that encompasses
all the checks, and a function that just matches the name. The former is
the property that should be used to decide whether to use the section
size, and the latter is used both as part of determining the value for
that field and for suppressing warnings about zero-length capabilities
(as would otherwise seen for empty sections).

This fixes the following assertion previously triggered by the new test
(copied from the existing CHERI capreloc-aether test):

  Assertion failed: (OffsetInOS <= OS->size), function getMorelloTargetSize, file lld/ELF/Arch/Cheri.cpp, line 739.

scalarizing vectors of capabilities.

There are many places where the frame pointer is hardcoded to x29 for
AArch64 (and respectively to c29 in a capability context). This includes
both internal variable/enum names and user-visible strings (e.g. "x29"
is considered an alternative name for "fp" rather than the other way
around).

This patch removes fp and cfp from all the register number enums and
from register infos and explicitly defines and uses x29/c29 instead in
the code. We preserve the behaviour that the name of the frame pointer
is "fp"/"cfp" rather than "xN" or "cN".

Note that since we no longer know the frame pointer statically, we need
to update the register infos with frame pointer information dynamically.
For this purpose, a helper function is added that will mark one of the
registers as the frame pointer or capability frame pointer. We call this
in every file that makes use of the register infos directly, usually in
some constructor or initialization function. One additional change is
necessary in GDBRemoteRegisterContext.cpp - this file was including
RegisterInfos_arm64.h inside a function, which made it impossible
to define the helper. We therefore had to move things one scope up.

This isn't a nice way to refactor this code, but I think it's the best
way to minimize merge pains in the future.

Switch triple from 'aarch64-none-elf' to 'aarch64-linux-android', since
that is our main target at the moment. The behaviour is subtly
different, so we need to update a test because the CFA now remains
relative to the SP even when the FP changes.

Darwin's ABI prepends an additional underscore to every C/C++ identifier
and llvm-cxxfilt by default will remove this. Thus ensure the underscore
doesn't get stripped in this test so that the symbols are properly
demangled regardless of the OS.

The latter is only supported in csh and zsh, but not bash.

a triple.

based on the IP.

This also fixes issues in LocalProgramCounter when the pointer
is not sealed.

without symbols.

These are executable capabilities, so their size won't depend
on the symbol size anyway.

1. DSO-internal function addresses are generated from PCC
and sealed. This avoids a load instruction and should
slightly reduce the number of relocations required.

2. Capability stubs for internal objects are merged into
a table to allow relative addressing. This allows sharing
the adrp instruction between multiple uses of the capability
table. The downside to this is that it can increase register
pressure.

3. We perform a post-legalization combine to push constant
pointer adds into adrp (this can remove add instructions).
This usually happens pre-legalization, but re-using that
machinery would require either changes to the SelectionDAG
builder or doing the load from the capability table in IR
(which would affect some other SelectionDAG transforms).

 - For CIE version 1 (e.g. in DWARF 2.0.0), the return_address_register
   field is a ubyte [0..255].

 - For CIE version 3 (e.g. in DWARF 3), the field is instead a ULEB128
   constant.

Previously, libunwind accepted a CIE version of 1 or 3, but always
parsed the field as ULEB128.

Clang always outputs CIE version 1 into .eh_frame. (It can output CIE
version 3 or 4, but only into .debug_frame.)

Differential Revision: https://reviews.llvm.org/D83741

__eh_frame_start
__eh_frame_end
__eh_frame_hdr_start
__eh_frame_hdr_end

These symbols are used by libunwind for baremetal to get the start
and end of the .eh_frame and .eh_frame_hdr sections. Make sure
these are not affected by any further section padding done by
--morello-c64-plt.

format

We need to forbid the pointer mode for integers in pure-capability C,
only permitting it on actual pointer types (which makes it rather
redundant, but we can support it so might as well appease those who
still think this attribute is a good idea) and preserving the fact that
it's a pointer rather than creating a 64-bit/128-bit plain integer. We
also need to do the same for unwind_word (which is a particularly bad
concept for split register files and for hard float ABIs). For
completeness, add a capability mode to complement the pointer mode.

Conversely, forbid integer modes for capabilities, and thus pointers in
pure-capability C, in order to avoid loss of provenance; if a plain
integer is desired, the code should be modified to use one, not a
pointer type.

Whilst here, make the implementation of getRegisterWidth clearer.
Currently PointerWidth is left as the address width for pure-capability
code which is confusing and should one day be fixed, likely by
introducing a separate PointerRange/AddressWidth. Outside of X86 (which
overrides this anyway), this is used solely for the word mode, and that
is likely intended to be a plain integer, not a capability, although
given that X86 makes it an AVX type perhaps we should make it a
capability if the intent is to use it for memcpy. This at least
preserves the current behaviour and can be revisited later.

Some of the tests that we added or touched weren't included in
check-lldb or any other build target after 99451b44. This commit
moves them in the appropriate new place.

Add support for end-to-end lldb tests on an FVP model running Android.

The tests live in lldb/morello-android-tests and can be run independently
of the llvm build system. Each test has its own subdirectory containing:
  * source files,
  * an Android.bp describing how to build the target for the test, and
  * a ${TESTNAME}.test file containing lldb commands and FileCheck
  expectations.

Note that we do not use RUN lines to drive the setup of the tests.
Instead, all the tests are expected to have the same setup, consisting
of the following steps:
  1. Compiling the target executable
  2. Uploading it to the model (with adb)
  3. Starting lldb-server on the model
  4. Starting lldb on the host, in batch mode, and checking its output

For more details check out the README.

This patch adds a single test to serve as an example.

in the pure capability ABI.

We previously didn't have a case statement for this.

unsupported with clang-7

because pc + offset is aligned down.

We don't have fenv.h and we can't run echo in a baremetal
environment.

Enable lldb shell tests to use a different python than the system one
(/usr/bin/python3). This makes it possible for us to point lldb to the
python it was compiled with, by setting the value of PYTHONHOME in the
environment when running the tests.

This partially reverts commit 8fdf909c.

This was no longer being added after a7e15b06

The "frame recognizer" command depends on LLDB_ENABLE_PYTHON.

It now prints the target address instead of a relative offset.
See commit 7f36cb1f and
aa708763

It now prints the target address instead of a relative offset.
See commit 7f36cb1f and
aa708763

This change is no longer required after the latest upstream CHERI-LLVM merge.

This reverts commit bd1c75c7.

The change for https://reviews.llvm.org/D90897 tests FirstSec/LastSec,
that was added in https://reviews.llvm.org/D90458, instead of Sections list.

Cherry-picking https://reviews.llvm.org/D90458 is non-trivial at this point,
so temporarily use Sections list in the test.

Revert after D90458 is merged.

Change-Id: Iad97ca3b76a4f6561823f4d9141f30e31e85ff68