The C-style cast was discarding const and casting to `(uint8_t *)` at
the same time, causing a warning. Add const and non-const versions of
the method.
BaseMem::setSize() should not be used anymore as the only memory
operand that understands size is x86::Mem, which makes it x86
specific.
The reason is that other architectures require more bits, so for
example arm::Mem uses the storage used by x86 size for storing
other information such as offset mode, and possibly more information
will be needed in the future to support AArch64 SVE or SME, etc...
At the moment BaseMem::setSize() has been deprecated, so code using
it would still compile, but with a warning. It will be removed in
the future though.
For some reason the growing strategy of asmjit::String was too
aggressive, basically reaching the maximum doubling capacity too
fast (after the first reallocation). This code adapts the current
vector growing strategy to be used also by asmjit::String, which
doubles the capacity until a threshold is reached and then grows
linearly.
During bin-packing, a single function nonOverlappingUnionOf() is
called many times to calculate a union of one live ranges with
another. Before this change it used ZoneVector::reserve() to make
sure that there is enough space for the union, however, in some
cases this is not ideal in case that the union grows every time
the function is called. In that case it's reallocating the vector
many times, which affects performance.
Instead of calling reserve(), a new function growingReserve() was
added to tell the vector to grow when it needs to reallocate.
In addition, this change fixes some documentation regarding the
use of JitAllocator (Explicit Code Relocation section in core.h).
This feature has been disabled for a long time so
it could be tested properly, but production didn't
reveal any issues.
When try mode is enabled the RA will try to allocate
the reassignment first to avoid possibly having to
emit code in a separate block (try mode basically
"tries" to emit code before a branch and not as a
consequence of it).
C++20 deprecates mixing enums of different types (comparisons, etc...),
however, we use enums instea of "static constexpr" in classes to define
constants, because otherwise we would have to give such constants
storage - this is required for up to C++14 and since we still support
C++11 we have to keep using enums...
The problem is that the rewriter must also rewrite an instruction
ID in case that it's a [K|V]MOV[B|W|D|Q] instruction that moves
from either K or SIMD register to GP register. when such instruction
is rewritten in a way that it ends up as "xMOVx GP, [MEM]" it would
be invalid if it's not changed to a general purpose MOV.
The problem can only happen in case that the compiler spills a
virtual register, which is then moved to a scalar register.
In addition, checks were added to MOVD|MOVQ to ensure that when an
invalid instruction is emitted it's not ignored as it used to be.
* clang-18 on now enabled on CI and used for static analysis
* return error when X86Internal_setupSaveRestoreInfo() is called
with invalid register group. Should never happen though.
Register allocator now tries to allocate preserved registers last,
improving prolog/epilog sequence especially on AArch64, which has
calling conventions that require to preserve both GP and vector
registers.
The code was fine, however, some compilers may be able to optimize
it and in some border cases the features returned would be all zero.
This prevents such behavior.
* Each architecture now provides r32() and r64() functions for
register casting
* Each architecture now provides v128() function for register
casting, returning just Vec to make writing cross platform
code easier
* Added some basic condition code abstractions so it can be used
interchangeably across architectures
* Added back unlicense to asmjit database (now it's dual licensed)
mach_vm_remap() allows to create a dual mapping without having to
use a file descriptor, which has to open a file or shm memory. The
problem is that the recent macos version started displaying a popup
message when such file is opened and that annoyed a lot of users.
Thus, the initial code-path is no longer used, and mach_vm_remap()
is used instead.
This change only applies for x86 macs. Apple silicon doesn't allow
dual mapping and instead uses MAP_JIT.
MSVC incorrectly auto-vectorizes a loop that is used in liveness
analysis. Due to this bug the result is wrong, which then affects
how registers are allocated. This workarounds a C++ compiler bug.
A new HardenedRuntimeFlags::kDualMapping flag has been introduced to
detect whether dual mapping is provided by the target platform. This
flag can be set even when hardened runtime is not enforced in cases,
in which dual mapping is not available.
This fixes running unit tests on Apple hardware where dual mapping
is not available, but MAP_JIT is (AArch64 hardware).
Additionally, this changeset fixes using -msse2 flag on non-x86
targets, where compiler don't mind "-msse2" flag, but warns about
it. This makes the build 100% clean.
The standard AArch64 calling convention specifies that the minimum
size of each argument passed via stack is 8 bytes. However, Apple
doesn't follow this convention and only aligns stack to 8 bytes
when an argument has 8 bytes or more. This makes it incompatible
with the standard calling convention as two consecutive 32-bit
integers would occupy only 8 bytes instead of 16 when passed via
stack.
The problem was that the sign/zero extension was not working with
TypeIds, but it was working with register IDs, when emitted. Thus,
even when TypeId was uint8, for example, the register could be
uint32, so the final instruction emitted was 'movzx eax, eax', for
example, which was wrong.
* Renamed all eq() methods to equals() (consistency) (ABI)
* Reorganized some X86 instructions in X86 database
* Properly detect RISC-V CPU at compile time (Environment)
* Removed CallConvId::kNone in favor of defaulting to kCDecl (ABI)
* CallConvId::kHost is now alias to CallConvId::kCDecl (ABI)
* Added FloatABI to Environment to disginguish between softfloat
and hardfloat
* Added more AArch64 CPU features and their detection (ABI)
* Because of CallConvId changes it's now possible to run
compiler tests on 32-bit ARM (fixes a bug in test cases)
* Added QEMU to CI build matrix to test different architectures
This is unfortunately API/ABI break, which cannot be deprecated
instead of removed as arm::Vec would collide with a64::Vec if present.
This is a preparetion for ARM32 tooling.
This changeset contains an updated instruction database that brings
ARM32 instructions for the first time. It also updates instruction
database tooling especially for ARM64, which will also be used by
ARM32 generator.
Additionally, new operan has been added, which represents a register
list as used by ARM32 instruction set.
Other minor changes are related to ARM - some stuff had to be moved
to a64 namespace from arm namespace as it's incompatible between
32-bit and 64-bit ISA.
* Added more clang compilers on CI (CI)
* Added memory sanitizer to build matrix (CI)
* Use problem matcher in all builds (CI)
* Fixed the use of some constructs in tests
* Fixed warnings about unused functions in tests
* Fixed warnings about unused variables caused by some build options
* Fixed tests to be clean with MSAN (zeroing memory filled by JIT code)
* Removed -Wclass-memaccess (gcc) from ignored warnings
* Removed -Wconstant-logical-operand (clang) from ignored warnings
* Removed -Wunnamed-type-template-args (clang) from ignored warnings
* Reworked InstData and InstExData to not cause UB (ABI break)
Unfortunately the existing InstData and InstExData was not good for static
analysis and in general compilers emitted warnings regarding accessing
InstNode::_opArray. The reason was that InstExNode added one or two
more operands which extended InstData::_opArray, but there was no way to
tell the C++ compiler about this layout.
It has been changed to InstNode having no operands and InstNodeWithOperands
being templatized for the right number of operands. Nodes that need to
inherit InstNode would just inherit InstNodeWithOperands<N>. It works the
same way as before, just the class hierarchy changed a little.
OSUtils::getTickCount() is a legacy function that used to have some
magic to actually return correct values. However, since the function
was not used by any open-source project on GH it's better if it's just
removed so it doesn't have to be maintained. It had no tests and since
there is <chrono> in C++ it really has no purpose anymore.
In addition, use 64-bit file API by default on some platforms as this
is what most applications use anyway. This is just a little optimization
to avoid using both open() and open64() functions in an application if
it already uses 64-bit file API.
* Added more cmake options to control features to be built
* Added a detection whether to link to libpthread (if not part of libc)
* Added a detection whether to link to librt (if not part of libc)
* Removed pure virtual functions as this collides with -nostdlib
* As a result the support for ASMJIT_NO_STDCXX has improved
We don't import either architecture in archtraits so include
environment.h to ensure Environment is visible, and don't use
__builtin__clear_cache in Emscripten because it's unsupported.
