LLVM 17.0.1 Release Notes

Introduction

This document contains the release notes for the LLVM Compiler Infrastructure, release 17.0.1. Here we describe the status of LLVM, including major improvements from the previous release, improvements in various subprojects of LLVM, and some of the current users of the code. All LLVM releases may be downloaded from the LLVM releases web site.

For more information about LLVM, including information about the latest release, please check out the main LLVM web site. If you have questions or comments, the Discourse forums is a good place to ask them.

Note that if you are reading this file from a Git checkout or the main LLVM web page, this document applies to the next release, not the current one. To see the release notes for a specific release, please see the releases page.

Non-comprehensive list of changes in this release

  • …

Update on required toolchains to build LLVM

With LLVM 17.x we raised the version requirement of CMake used to build LLVM. The new requirements are as follows:

  • CMake >= 3.20.0

Changes to the LLVM IR

  • Typed pointers are no longer supported and the -opaque-pointers option has been removed. See the opaque pointers documentation for migration instructions.

  • The nofpclass attribute was introduced. This allows more optimizations around special floating point value comparisons.

  • Introduced new llvm.ldexp and llvm.experimental.constrained.ldexp intrinsics.

  • Introduced new llvm.frexp intrinsic.

  • The constant expression variants of the following instructions have been removed:

    • select

  • Introduced a set of experimental convergence control intrinsics to explicitly define the semantics of convergent operations.

Changes to LLVM infrastructure

  • The legacy optimization pipeline (PassManagerBuilder.h) has been removed. See the new pass manager docs for how to use the new pass manager APIs.

  • Alloca merging in the inliner has been removed, since it only worked with the legacy inliner pass. Backend stack coloring should handle cases alloca merging initially set out to handle.

  • InstructionSimplify APIs now require instructions be inserted into a parent function.

  • A new FatLTO pipeline was added to support generating object files that have both machine code and LTO compatible bitcode. See the FatLTO documentation and the original RFC for more details.

Changes to building LLVM

Changes to TableGen

  • Named arguments are supported. Arguments can be specified in the form of name=value.

Changes to Optimizations

  • llvm.assume now recognizes certain floating-point tests. e.g. __builtin_assume(!isnan(x))

Changes to the AArch64 Backend

  • Added Assembly Support for the 2022 A-profile extensions FEAT_GCS (Guarded Control Stacks), FEAT_CHK (Check Feature Status), and FEAT_ATS1A.

  • Support for preserve_all calling convention is added.

  • Added support for missing arch extensions in the assembly directives .arch <level>+<ext> and .arch_extension.

  • Fixed handling of .arch <level> in assembly, without using any +<ext> suffix. Previously this had no effect at all if no extensions were supplied. Now .arch <level> can be used to enable all the extensions that are included in a higher level than what is specified on the command line, or for disabling unwanted extensions if setting it to a lower level. This fixes PR32873 <https://github.com/llvm/llvm-project/issues/32220>.

Changes to the AMDGPU Backend

  • More fine-grained synchronization around barriers for newer architectures (gfx90a+, gfx10+). The AMDGPU backend now omits previously automatically generated waitcnt instructions before barriers, allowing for more precise control. Users must now use memory fences to implement fine-grained synchronization strategies around barriers. Refer to AMDGPU memory model.

  • Address space 7, used for buffer fat pointers has been added. It is non-integral and has 160-bit pointers (a 128-bit raw buffer resource and a 32-bit offset) and 32-bit indices. This is part of ongoing work to improve the usability of buffer operations. Refer to AMDGPU address spaces.

  • Address space 8, used for buffer resources has been added. It is non-integral and has 128-bit pointers, which correspond to buffer resources in the underlying hardware. These pointers should not be used with getelementptr or other LLVM memory instructions, and can be created with the llvm.amdgcn.make.buffer.rsrc intrinsic. Refer to AMDGPU address spaces.

  • New versions of the intrinsics for working with buffer resources have been added. These llvm.amdgcn.*.ptr.[t]buffer.* intrinsics have the same semantics as the old llvm.amdgcn.*.[t]buffer.* intrinsics, except that their rsrc arguments are represented by a ptr addrspace(8) instead of a <4 x i32>. This improves the interaction between AMDGPU buffer operations and the LLVM memory model, and so the non .ptr intrinsics are deprecated.

  • Backend now performs range merging of “amdgpu-waves-per-eu” attribute based on known callers.

  • Certain atomicrmw operations are now optimized by performing a wave reduction if the access is uniform by default.

  • Removed llvm.amdgcn.atomic.inc and llvm.amdgcn.atomic.dec intrinsics. atomicrmw should be used instead with uinc_wrap and udec_wrap.

  • Added llvm.amdgcn.log.f32 intrinsic. This provides direct access to v_log_f32.

  • Added llvm.amdgcn.exp2.f32 intrinsic. This provides direct access to v_exp_f32.

  • llvm.log2.f32, llvm.log10.f32, and llvm.log.f32 are now lowered accurately. Use llvm.amdgcn.log.f32 to access the old behavior for llvm.log2.f32.

  • llvm.exp2.f32 and llvm.exp.f32 are now lowered accurately. Use llvm.amdgcn.exp2.f32 to access the old behavior for llvm.exp2.f32.

  • Implemented new 1ulp IEEE lowering strategy for float reciprocal which saves 2 instructions. This is used by default for OpenCL on gfx9+. With contract flags, this will fold into a 1 ulp rsqrt.

  • Implemented new 2ulp IEEE lowering strategy for float reciprocal. This is used by default for OpenCL on gfx9+.

  • llvm.sqrt.f64 is now lowered correctly. Use llvm.amdgcn.sqrt.f64 for raw instruction access.

  • Deprecate llvm.amdgcn.ldexp intrinsic. llvm.ldexp should be used instead.

Changes to the ARM Backend

  • The hard-float ABI is now available in Armv8.1-M configurations that have integer MVE instructions (and therefore have FP registers) but no scalar or vector floating point computation.

  • The .arm directive now aligns code to the next 4-byte boundary, and the .thumb directive aligns code to the next 2-byte boundary.

Changes to the AVR Backend

  • …

Changes to the DirectX Backend

Changes to the Hexagon Backend

  • …

Changes to the LoongArch Backend

  • Adds assembler/disassembler support for the LSX, LASX, LVZ and LBT ISA extensions.

  • The lp64s ABI is supported now and has been tested on Rust bare-matal target.

  • A target feature ual is introduced to allow unaligned memory accesses and this feature is enabled by default for generic 64-bit processors.

  • Adds support for the large code model, which is equivalent to GCC’s extreme one.

  • Assorted codegen improvements.

  • llvm-objcopy now supports processing LoongArch objects.

Changes to the MIPS Backend

  • …

Changes to the PowerPC Backend

  • Improved code sequence of materializing 64-bit immediate numbers, expanding is.fpclass intrinsic and forwarding stores.

  • Implemented DFP instructions (for use via inline asm).

  • Improved code gen for vector add.

  • Added ability to show statistics of number of entries in the TOC.

  • Added Binary Coded Decimal Assist instructions (for use via inline asm).

  • Added basic support for vector functions in GlobalISel.

  • Added additional X-Form load and store instruction generation for TLS accesses.

  • PPC64LE backend is added to JITLink.

  • Added various bug fixes and optimizations.

  • Added function pointer alignment to the DataLayout for Power, which lets us make more informed choices about what this alignment defaults to for various purposes (e.g., C++ pointers to member). If the target ABI uses function descriptor objects, this is the alignment we would emit the descriptor with. Otherwise, a function pointer points to a global entry point, so this is at least the alignment for code on Power (i.e., 4-bytes).

AIX Support/improvements:

  • A new option -mxcoff-roptr is added to clang and llc. When this option is present, constant objects with relocatable address values are put into the RO data section. This option should be used with the -fdata-sections option, and is not supported with -fno-data-sections.

  • Taught the profile runtime to check for a build-id string. Build-id strings can be created via the -mxcoff-build-id option.

  • Removed -ppc-quadword-atomics which only affected lock-free quadword atomics on AIX. Now backend generates lock-free quadword atomics code on AIX by default. To support lock-free quadword atomics in libatomic, the OS level must be at least AIX 7.2 TL5 SP3 with libc++.rte of version 17.1.1 or above installed.

  • Integrated assembler is enabled by default on AIX.

  • System assembler is always used to compile assembly files on AIX.

  • Added support for local-exec TLS.

  • Added a new option, --traceback-table, to llvm-objdump to print out the traceback table information for XCOFF object files.

  • Added llvm-ar object mode options -X32, -X64, -X32-64, and -Xany.

  • Changed the default name of the text-section csect to be an empty string instead of .text. This change does not affect the behaviour of the program.

  • Fixed a problem when the personality routine for the legacy AIX xlclang++ compiler uses the stack slot to pass the exception object to the landing pad. Runtime routine __xlc_exception_handle() invoked by the landing pad to retrieve the exception object now skips frames not associated with functions that are C++ EH-aware because the compiler sometimes generates a wrapper of __xlc_exception_handle() for optimization purposes.

Changes to the RISC-V Backend

  • Assembler support for version 1.0.1 of the Zcb extension was added.

  • Zca, Zcf, and Zcd extensions were upgraded to version 1.0.1.

  • vsetvli intrinsics no longer have side effects. They may now be combined, moved, deleted, etc. by optimizations.

  • Adds support for the vendor-defined XTHeadBa (address-generation) extension.

  • Adds support for the vendor-defined XTHeadBb (basic bit-manipulation) extension.

  • Adds support for the vendor-defined XTHeadBs (single-bit) extension.

  • Adds support for the vendor-defined XTHeadCondMov (conditional move) extension.

  • Adds support for the vendor-defined XTHeadMac (multiply-accumulate instructions) extension.

  • Added support for the vendor-defined XTHeadMemPair (two-GPR memory operations) extension disassembler/assembler.

  • Added support for the vendor-defined XTHeadMemIdx (indexed memory operations) extension disassembler/assembler.

  • Added support for the vendor-defined Xsfvcp (SiFive VCIX) extension disassembler/assembler.

  • Added support for the vendor-defined Xsfcie (SiFive CIE) extension disassembler/assembler.

  • Support for the now-ratified Zawrs extension is no longer experimental.

  • Adds support for the vendor-defined XTHeadCmo (cache management operations) extension.

  • Adds support for the vendor-defined XTHeadSync (multi-core synchronization instructions) extension.

  • Added support for the vendor-defined XTHeadFMemIdx (indexed memory operations for floating point) extension.

  • Assembler support for RV64E was added.

  • Assembler support was added for the experimental Zicond (integer conditional operations) extension.

  • I, F, D, and A extension versions have been update to the 20191214 spec versions. New version I2.1, F2.2, D2.2, A2.1. This should not impact code generation. Immpacts versions accepted in -march and reported in ELF attributes.

  • Changed the ShadowCallStack register from x18 (s2) to x3 (gp). Note this breaks the existing non-standard ABI for ShadowCallStack on RISC-V, but conforms with the new “platform register” defined in the RISC-V psABI (for more details see the psABI discussion).

  • Added support for Zfa extension version 0.2.

  • Updated support experimental vector crypto extensions to version 0.5.1 of the specification.

  • Removed N extension (User-Level Interrupts) CSR names in the assembler.

  • RISCV::parseCPUKind and RISCV::checkCPUKind were merged into a single RISCV::parseCPU. The CPUKind enum is no longer part of the RISCVTargetParser.h interface. Similar for parseTuneCPUkind and checkTuneCPUKind.

  • Add sifive-x280 processor.

  • Zve32f is no longer allowed with Zfinx. Zve64d is no longer allowed with Zdinx.

  • Assembly support was added for the experimental Zfbfmin (scalar BF16 conversions), Zvfbfmin (vector BF16 conversions), and Zvfbfwma (vector BF16 widening mul-add) extensions.

  • Added assembler/disassembler support for the experimental Zacas (atomic compare-and-swap) extension.

  • Zvfh extension version was upgraded to 1.0 and is no longer experimental.

Changes to the WebAssembly Backend

  • Function annotations (__attribute__((annotate(<name>)))) now generate custom sections in the Wasm output file. A custom section for each unique name will be created that contains each function index the annotation applies to.

Changes to the Windows Target

Changes to the X86 Backend

  • __builtin_unpredictable (unpredictable metadata in LLVM IR), is handled by X86 Backend. X86CmovConversion pass now respects this builtin and does not convert CMOVs to branches.

  • Add support for the PBNDKB instruction.

  • Support ISA of SHA512.

  • Support ISA of SM3.

  • Support ISA of SM4.

  • Support ISA of AVX-VNNI-INT16.

  • -mcpu=graniterapids-d is now supported.

Changes to the OCaml bindings

Changes to the Python bindings

  • The python bindings have been removed.

Changes to the C API

  • LLVMContextSetOpaquePointers, a temporary API to pin to legacy typed pointer, has been removed.

  • Functions for adding legacy passes like LLVMAddInstructionCombiningPass have been removed.

  • Removed LLVMPassManagerBuilderRef and functions interacting with it. These belonged to the no longer supported legacy pass manager.

  • Functions for initializing legacy passes like LLVMInitializeInstCombine have been removed. Calls to such functions can simply be dropped, as they are no longer necessary.

  • LLVMPassRegistryRef and LLVMGetGlobalPassRegistry, which were only useful in conjunction with initialization functions, have been removed.

  • As part of the opaque pointer transition, LLVMGetElementType no longer gives the pointee type of a pointer type.

  • The following functions for creating constant expressions have been removed, because the underlying constant expressions are no longer supported. Instead, an instruction should be created using the LLVMBuildXYZ APIs, which will constant fold the operands if possible and create an instruction otherwise:

    • LLVMConstSelect

Changes to the CodeGen infrastructure

  • llvm.memcpy, llvm.memmove and llvm.memset are now expanded into loops by default for targets which do not report the corresponding library function is available.

Changes to the Metadata Info

Changes to the Debug Info

  • The DWARFv5 feature of attaching DW_AT_default_value to defaulted template parameters will now be available in any non-strict DWARF mode and in a wider range of cases than previously. (D139953, D139988)

  • The DW_AT_name on DW_AT_typedefs for alias templates will now omit defaulted template parameters. (D142268)

  • The experimental @llvm.dbg.addr intrinsic has been removed (D144801). IR inputs with this intrinsic are auto-upgraded to @llvm.dbg.value with DW_OP_deref appended to the DIExpression (D144793).

  • When a template class annotated with the [[clang::preferred_name]] attribute were to appear in a DW_AT_type, the type will now be that of the preferred_name instead. This change is only enabled when compiling with -glldb. (D145803)

Changes to the LLVM tools

  • llvm-lib now supports the /def option for generating a Windows import library from a definition file.

  • Made significant changes to JSON output format of llvm-readobj/llvm-readelf to improve correctness and clarity.

Changes to LLDB

  • In the results of commands such as expr and frame var, type summaries will now omit defaulted template parameters. The full template parameter list can still be viewed with expr --raw-output/frame var --raw-output. (D141828)

  • LLDB is now able to show the subtype of signals found in a core file. For example memory tagging specific segfaults such as SIGSEGV: sync tag check fault.

  • LLDB can now display register fields if they are described in target XML sent by a debug server such as gdbserver (lldb-server does not currently produce this information). Fields are only printed when reading named registers, for example register read cpsr. They are not shown when reading a register set, register read -s 0.

  • A new command register info was added. This command will tell you everything that LLDB knows about a register. Based on what LLDB already knows and what the debug server tells it. Including but not limited to, the size, where it is read from and the fields that the register contains.

  • AArch64 Linux targets now provide access to the Thread Local Storage register tpidr.

Changes to Sanitizers

  • For Darwin users that override weak symbols, note that the dynamic linker will only consider symbols in other mach-o modules which themselves contain at least one weak symbol. A consequence is that if your program or dylib contains an intended override of a weak symbol, then it must contain at least one weak symbol as well for the override to take effect.

    Example:

    // Add this to make sure your override takes effect
__attribute__((weak,unused)) unsigned __enableOverrides;

// Example override
extern "C" const char *__asan_default_options() { ... }

Changes to BOLT

  • Initial RISC-V (RV64GC) target support was added.

  • DWARFRewriter got new mechanism for more flexible handling of debug information. It raises debug information to IR level before performing updates, and IR is written out to the binary after updates are applied.

  • Stale profile matching was added under a flag –infer-stale-profile. It requires the use of a YAML profile, produced by perf2bolt using -w flag, or with –profile-format=yaml.

Other Changes

  • llvm::demangle now takes a std::string_view rather than a const std::string&. Be careful passing temporaries into llvm::demangle that don’t outlive the expression using llvm::demangle.

External Open Source Projects Using LLVM 15

  • A project…

Additional Information

A wide variety of additional information is available on the LLVM web page, in particular in the documentation section. The web page also contains versions of the API documentation which is up-to-date with the Git version of the source code. You can access versions of these documents specific to this release by going into the llvm/docs/ directory in the LLVM tree.

If you have any questions or comments about LLVM, please feel free to contact us via the Discourse forums.