* Renamed round to round_even
* Added round_half_up intrinsic
* Added floating-point mod
* Added a scalar version of floating-point abs and neg
* Added a behavior enum to specify how float to int conversion
handles out-of-range and NaN cases
* Updated some APX stuff in instruction database
* Inspired by the one of the pull requests (#483)
* Use std::make_unique to make a unique pointer (tests only)
* Use override instead of virtual to override a virtual member
function
* The bug was introduced by the latest changes. The rewriter
only interated basic blocks, but not possibly injected code
that converts one register state into another via a small
code sequence injected at the end of the function.
* Resolved some static analysis problems. The reported
use-after-free was a false positive constructing by
contradicting conditions - so new asserts guard this.
* Refactored the whole codebase to use snake_case convention to
name functions and variables, including member variables.
Class naming is unchanged and each starts with upper-case
character. The intention of this change is to make the source
code more readable and consistent across multiple projects
where AsmJit is currently used.
* Refactored support.h to make it more shareable across projects.
* x86::Vec now inherits from UniVec
* minor changes in JitAllocator and WriteScope in order to make
the size of WriteScope smaller
* added ZoneStatistics and Zone::statistics() getter
* improved x86::EmitHelper to use tables instead of choose() and
other mechanisms to pick between SSE and AVX instructions
* Refactored the whole codebase to use snake_case convention for
for functions names, function parameter names, struct members,
and variables
* Added a non-owning asmjit::Span<T> type and use into public API
to hide the usage of ZoneVector in CodeHolder, Builder, and
Compiler. Users now only get Span (with data and size), which
doesn't require users to know about ZoneVector
* Removed RAWorkId from RATiedReg in favor of RAWorkReg*
* Removed GEN from LiveInfo as it's not needed by CFG construction
to save memory (GEN was merged with LIVE-IN bits). The remaining
LIVE-IN, LIVE-OUT, and KILL bits are enough, however KILL bits may
be removed in the future as KILL bits are not needed after LIVE-IN
and LIVE-OUT converged
* Optimized the representation of LIVE-IN, LIVE-OUT, and KILL bits
per block. Now only registers that live across multiple basic
blocks are included here, which means that virtual registers that
only live in a single block are not included and won't be overhead
during liveness analysis. This optimization alone can make liveness
analysis 90% faster depending on the code generated (more virtual
registers that only live in a single basic block -> more gains)
* Optimized building liveness information bits per block. The new
code uses an optimized algorithm to prevent too many traversals
and uses a more optimized code for a case in which not too many
registers are used (it avoids array operations if the number of
all virtual registers within the function fits a single BitWord)
* Optimized code that computes which virtual register is only used
in a single basic block - this aims to optimize register allocator
in the future by using a designed code path for allocating regs
only used in a single basic block
* Reduced the information required for each live-span, which is used
by bin-packing. Now the struct is 8 bytes, which is good for a lot
of optimizations C++ compiler can do
* Added UniCompiler (ujit) which can be used to share code paths
between X86, X86_64, and AArch64 code generation (experimental).
* Added first node to Zone so the reset is simpler
* Added x86::Xmm/Ymm/Zmm deprecated aliases of x86::Vec
to make user code not break when using these deprecated
types
* Documentation fixes and clarifications
* Reworked register operands - all vector registers are now
platform::Vec deriving from UniVec (universal vector operand),
additionally, there is no platform::Reg, instead asmjit::Reg
provides all necessary features to make it a base register for
each target architecture
* Reworked casting between registers - now architecture agnostic
names are preferred - use Gp32 instead of Gpd or GpW, Gp64
instead of Gpq and GpX, etc...
* Reworked vector registers and their names - architecture
agnostic naming is now preferred Vec32, Vec64, Vec128, etc...
* Reworked naming conventions used across AsmJit - for clarity
Identifiers are now prefixed with the type, like sectionId(),
labelId(), etc...
* Reworked how Zone and ZoneAllocator are used across AsmJit,
prefering Zone in most cases and ZoneAllocator only for
containers - this change alone achieves around 5% better
performance of Builder and Compiler
* Reworked LabelEntry - decreased the size of the base entry
to 16 bytes for anonymous and unnamed labels. Avoided an
indirection when using labelEntries() - LabelEntry is now
a value and not a pointer
* Renamed LabelLink to Fixup
* Added a new header <asmjit/host.h> which would include
<asmjit/core.h> + target tools for the host architecture,
if enabled and supported
* Added new AArch64 instructions (BTI, CSSC, CHKFEAT)
* Added a mvn_ alternative of mvn instruction (fix for Windows
ARM64 SDK)
* Added more AArch64 CPU features to CpuInfo
* Added better support for Apple CPU detection (Apple M3, M4)
* Added a new benchmarking tool asmjit_bench_overhead, which
benchmarks the overhead of CodeHolder::init()/reset() and
creating/attaching emitters to it. Thanks to the benchmark the
most common code-paths were optimized
* Added a new benchmarking tool asmjit_bench_regalloc, which
aims to benchmark the cost and complexity of register allocation.
* Renamed asmjit_test_perf to asmjit_bench_codegen to make it
clear what is a test and what is a benchmark
The purpose of this change is to minimize casting when using Vec
as a type for all vector registers in user code. In these cases
AsmJit required to cast Vec to Xmm to call emitter methods. After
this change it's no longer necessary to do that, which simplifies
user code.
* Instructions wr[u]ss[d|q] no longer accept register as the first
operand (that was a bug to accept this form)
* Moved APX version of legacy instructions closer so they are next
to each other
* Removed AVX512_ER, AVX512_PF, AVX512_4FMAPS, and AVX512_4VNNIW
extensions and corresponding instructions (these were never
advertised by any x86 CPU and were only used by Xeon Phi acc.,
which AsmJit never supported)
* Removed CPU extensions HLE, MPX, and TSX
* Kept extension RTM, which is only for backward compatibility to
recognize instructions, but it's no longer checked by CpuInfo as
it's been deprecated together with HLE and MPX
* The xtest instruction now reports it requires RTM
* Reorganized x86 extensions a bit - they are now reordered to group
them by category, preparing for the future where extension IDs will
be always added after existing records for ABI compatibility
* Instruction vcvtneps2bf16 no longer accepts form without an explicit
memory operand size
* Removed aliased instructions in CMOVcc, Jcc, And SETcc categories,
now there is only a single instruction id for all aliased instructions.
* Added a new feature to always show instruction aliases in Logger, which
includes formatting instructio nodes (Builder, Compiler)
Instruction DB-only updates (not applied to C++ yet):
* AsmJit DB from now uses the same license as AsmJit (Zlib) and
no longer applies dual licensing (Zlib and Public Domain)
* Added support for aggregated instruction definitions in
x86 instruction database, which should simplify the maintenance
and reduce bugs (also the syntax is comparable to descriptions
used by Intel APX instruction manuals)
* Added support for APX instructions and new features
* Added support for AVX10.1 and AVX10.2 instructions (both new
instructions and new encodings of existing instructions)
* Added support for MOVRS instructions
* Added support for KL instructions (loadiwkey)
* Added support for AESKLE instructions
* Added support for AESKLEWIDE_KL instructions
* Added support for AMX_[AVX512|MOVRS|FP8|TF32|TRANSPOSE]
* NOTE: None of the instruction additions is currently used by
Asmjit, it's a pure database update that needs more work to
make all the instructions available in future AsmJit
* The implementation tries to detect whether a virtual register
only lives in a single basic block and then uses a move approach
instead of spill/alloc when reallocating
* Additionally, the implementation now improves the use of scratch
registers during function arguments allocation - scratch is only
reserved when it's actually needed
Changes the reassognment decision used by local register allocator.
When the virtual register is killed by the instruction (or has a
separate output slot) it would be reassigned instead of spilled.
This should be a minor improvement for certain border cases.
* Fixes a potential UB in number to string conversion because of
a possible undefined behavior caused by -int64_t(a) code. The
fix replaces the code with Support::neg() function, which was
designed for exactly this.
* Little documentation fixes
* Added virtual destructor to emit helpers to silence warnings of
some compilers (it's totally useless change though as it changes
nothing in reality - emit helpers are allocated mostly on stack)
* The first operand (destination) is read/write and not overwrite
In addition, added the following new AArch64 instructions to DB:
* CPA extensions (DB-only)
* FAMINMAX extensions (DB-only)
* FP8 ASIMD extensions (DB-only)
* LUT extensions (DB-only)
Extend option in ADD, ADDS, SUB, SUBS, CMP, and CMN instructions
doesn't always use the same second register type. For example when
extending from a BYTE the second source register must be W and not
X.
This change makes sure that the assembler accepts the correct
combination and refuses the incorrect one.
IMPORTANT: Although this is not an ABI change, the new behavior
can break AArch64 code that used the incorrect signatures.
* The latest cmake versions started showing warnings about the
minimum version supported as there are possibly some breaking
changes not affecting us
* Reworked some bits in CMakeLists.txt to take advantage of the
raised version
* Removed the use of policies that are now enabled by cmake by
default
* Removed deprecated build options
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...
