LLVM 11.0.0 Release Notes

Introduction

This document contains the release notes for the LLVM Compiler Infrastructure, release 11.0.0. 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 LLVM Developer's Mailing List is a good place to send them.

Deprecated and Removed Features/APIs

• BG/Q support, including QPX, will be removed in the 12.0.0 release.

Non-comprehensive list of changes in this release

• The llgo frontend has been removed for now, but may be resurrected in the future.

Changes to the LLVM IR

• The callsite attribute vector-function-abi-variant has been added to describe the mapping between scalar functions and vector functions, to enable vectorization of call sites. The information provided by the attribute is interfaced via the API provided by the VFDatabase class. When scanning through the set of vector functions associated with a scalar call, the loop vectorizer now relies on VFDatabase, instead of TargetLibraryInfo.
• dereferenceable attributes and metadata on pointers no longer imply anything about the alignment of the pointer in question. Previously, some optimizations would make assumptions based on the type of the pointer. This behavior was undocumented. To preserve optimizations, frontends may need to be updated to generate appropriate align attributes and metadata.
• The DIModule metadata is extended to contain file and line number information. This information is used to represent Fortran modules debug info at IR level.
• LLVM IR now supports two distinct llvm::FixedVectorType and llvm::ScalableVectorType vector types, both derived from the base class llvm::VectorType. A number of algorithms dealing with IR vector types have been updated to make sure they work for both scalable and fixed vector types. Where possible, the code has been made generic to cover both cases using the base class. Specifically, places that were using the type unsigned to count the number of lanes of a vector are now using llvm::ElementCount. In places where uint64_t was used to denote the size in bits of a IR type we have partially migrated the codebase to using llvm::TypeSize.
• Branching on undef/poison is undefined behavior. It is needed for correctly analyzing value ranges based on branch conditions. This is consistent with MSan's behavior as well.
• memset/memcpy/memmove can take undef/poison pointer(s) if the size to fill is zero.
• Passing undef/poison to a standard I/O library function call (printf/fputc/...) is undefined behavior. The new noundef attribute is attached to the functions' arguments. The full list is available at llvm::inferLibFuncAttributes.

Changes to building LLVM

• The LLVM project has started the migration towards Python 3, and the build system now prefers Python 3 whenever available. If the Python 3 interpreter (or libraries) are not found, the build system will, for the time being, fall back to Python 2. It is recommended that downstream projects migrate to Python 3 as Python 2 has been end-of-life'd by the Python Software Foundation.

Changes to the JIT infrastructure

• LLJIT now supports execution of static inits / deinits via the LLJIT::initialize and LLJIT::deinitialize methods
• Static libraries can now be added to a JITDylib using the StaticLibraryDefinitionGenerator class
• A C API has been added for OrcV2 (llvm-project/llvm/include/llvm-c/Orc.h)
• Several OrcV2 example projects have been added to llvm-project/llvm/examples/OrcV2Examples
• Many bug fixes and API improvements

Changes to the AArch64 Backend

• Back up and restore x18 in functions with windows calling convention on non-windows OSes.

• Clearly error out on unsupported relocations when targeting COFF, instead of silently accepting some (without being able to do what was requested).

• Implemented codegen support for the SVE C-language intrinsics documented in Arm C Language Extensions (ACLE) for SVE (version 00bet5). For more information, see the clang 11 release notes.

• Added support for Armv8.6-A:

  Assembly support for the following extensions:

  • Enhanced Counter Virtualization (ARMv8.6-ECV).
  • Fine Grained Traps (ARMv8.6-FGT).
  • Activity Monitors virtualization (ARMv8.6-AMU).
  • Data gathering hint (ARMv8.0-DGH).

  Assembly and intrinsics support for the Armv8.6-A Matrix Multiply extension for Neon and SVE vectors.

  Support for the ARMv8.2-BF16 BFloat16 extension. This includes a new C-level storage-only __bf16 type, a BFloat IR type, a bf16 MVT, and assembly and intrinsics support.

• Added support for Cortex-A34, Cortex-A77, Cortex-A78 and Cortex-X1 cores.

Changes to the ARM Backend

• Implemented C-language intrinsics for the full Arm v8.1-M MVE instruction set. <arm_mve.h> now supports the complete API defined in the Arm C Language Extensions.

• Added support for assembly for the optional Custom Datapath Extension (CDE) for Arm M-profile targets.

• Implemented C-language intrinsics <arm_cde.h> for the CDE instruction set.

• Clang now defaults to -fomit-frame-pointer when targeting non-Android Linux for arm and thumb when optimizations are enabled. Users that were previously not specifying a value and relying on the implicit compiler default may wish to specify -fno-omit-frame-pointer to get the old behavior. This improves compatibility with GCC.

• Added support for Armv8.6-A:

  Assembly and intrinsics support for the Armv8.6-A Matrix Multiply extension for Neon vectors.

  Support for the ARMv8.2-AA32BF16 BFloat16 extension. This includes a new C-level storage-only __bf16 type, a BFloat IR type, a bf16 MVT, and assembly and intrinsics support.

• Added support for CMSE.

• Added support for Cortex-M55, Cortex-A77, Cortex-A78 and Cortex-X1 cores.

• The Machine Outliner is now supported for ARM and Thumb2, it is not turned on by default and can be enabled with the -moutline clang flag.

Changes to the PowerPC Target

Optimization:

• Improved Loop Unroll-and-Jam legality checks, allowing it to handle more than two level loop nests
• Improved Loop Unroll to be able to unroll more loops
• Implemented an option to allow loop fusion to work on loops with different constant trip counts

Codegen:

• POWER10 support
  • Added PC Relative addressing
  • Added __int128 vector bool support
• Security enhancement via probe-stack attribute support to protect against stack clash
• Floating point support enhancements
  • Improved half precision and quad precision support, including GLIBC
  • constrained FP operation support for arithmetic/rounding/max/min
  • cleaning up fast math flags checks in DAGCombine, Legalizer, and Lowering
• Performance improvements from instruction exploitation, especially for vector permute on LE
• Scheduling enhancements
  • Added MacroFusion for POWER8
  • Added post-ra heuristics for POWER9
• Target dependent passes tuning
  • Updated LoopStrengthReduce to use instruction number as first priority
  • Enhanced MachineCombiner to expose more ILP
• Code quality and maintenance enhancements
  • Enabled more machine verification passes
  • Added ability to parse and emit additional extended mnemonics
  • Numerous bug fixes

AIX Support Improvements:

• Enabled compile and link such that a simple <stdio.h> "Hello World" program works with standard headers
• Added support for the C calling convention for non-vector code
• Implemented correct stack frame layout for functions
• In llvm-objdump, added support for relocations, improved selection of symbol labels, and added the --symbol-description option

Changes to the RISC-V Target

New features:

• After consultation through an RFC, the RISC-V backend now accepts patches for proposed instruction set extensions that have not yet been ratified. For these experimental extensions, there is no expectation of ongoing support - the compiler support will continue to change until the specification is finalised. In line with this policy, MC layer and code generation support was added for version 0.92 of the proposed Bit Manipulation Extension and MC layer support was added for version 0.8 of the proposed RISC-V Vector instruction set extension. As these extensions are not yet ratified, compiler support will continue to change to match the specifications until they are finalised.
• ELF attribute sections are now created, encoding information such as the ISA string.
• Support for saving/restoring callee-saved registers via libcalls (a code size optimisation).
• llvm-objdump will now print branch targets as part of disassembly.

Improvements:

• If an immediate can be generated using a pair of addi instructions, that pair will be selected rather than materialising the immediate into a separate register with an lui and addi pair.
• Multiplication by a constant was optimised.
• addi instructions are now folded into the offset of a load/store instruction even if the load/store itself has a non-zero offset, when it is safe to do so.
• Additional target hooks were implemented to minimise generation of unnecessary control flow instruction.
• The RISC-V backend's load/store peephole optimisation pass now supports constant pools, improving code generation for floating point constants.
• Debug scratch register names dscratch0 and dscratch1 are now recognised in addition to the legacy dscratch register name.
• Codegen for checking isnan was improved, removing a redundant and.
• The dret instruction is now supported by the MC layer.
• .option pic and .option nopic are now supported in assembly and .reloc was extended to support arbitrary relocation types.
• Scheduling info metadata was improved.
• The jump pseudo instruction is now supported.

Bug fixes:

• A failure to insert indirect branches in position independent code was fixed.
• The calculated expanded size of atomic pseudo operations was fixed, avoiding "fixup value out of range" errors during branch relaxation for some inputs.
• The mcountinhibit CSR is now recognised.
• The correct libcall is now emitted for converting a float/double to a 32-bit signed or unsigned integer on RV64 targets lacking the F or D extensions.

Changes to the SystemZ Target

• Added support for the MemorySanitizer and the LeakSanitizer.
• Added support for the -fstack-clash-protection command line option.
• Enhanced the assembler parser to allow using %r0 even in an address register context, and to allow specifying registers using plain integer numbers instead of register names everywhere.
• Fixed wrong code generation violating the platform ABI when passing a C++ class (not struct) type having only a single member of floating-point type.
• Fixed wrong code generation when using the vec_store_len_r or vec_load_len_r intrinsics with an immediate length argument of 16 or larger.
• Miscellaneous codegen enhancements, in particular to improve vector code.

Changes to the X86 Target

• Functions with the probe-stack attribute set to "inline-asm" are now protected against stack clash without the need of a third-party probing function and with limited impact on performance.
• -x86-enable-old-knl-abi command line switch has been removed. v32i16/v64i8 vectors are always passed in ZMM register when avx512f is enabled and avx512bw is disabled.
• Vectors larger than 512 bits with i16 or i8 elements will be passed in multiple ZMM registers when avx512f is enabled. Previously this required avx512bw otherwise they would split into multiple YMM registers. This means vXi16/vXi8 vectors are consistently treated the same as vXi32/vXi64/vXf64/vXf32 vectors of the same total width.
• Support was added for Intel AMX instructions.
• Support was added for TSXLDTRK instructions.
• A pass was added for mitigating the Load Value Injection vulnerability.
• The Speculative Execution Side Effect Suppression pass was added which can be used to as a last resort mitigation for speculative execution related CPU vulnerabilities.
• Improved recognition of boolean vector reductions with better MOVMSKB/PTEST handling
• Exteded recognition of rotation patterns to handle funnel shift as well, allowing us to remove the existing x86-specific SHLD/SHRD combine.

Changes to the AMDGPU Target

• The backend default denormal handling mode has been switched to on for all targets for all compute function types. Frontends wishing to retain the old behavior should explicitly request f32 denormal flushing.

Changes to the AVR Target

• Moved from an experimental backend to an official backend. AVR support is now included by default in all LLVM builds and releases and is available under the "avr-unknown-unknown" target triple.

Changes to the WebAssembly Target

• Programs which don't have a "main" function, called "reactors" are now properly supported, with a new -mexec-model=reactor flag. Programs which previously used -Wl,--no-entry to avoid having a main function should switch to this new flag, so that static initialization is properly performed.
• __attribute__((visibility("protected"))) now evokes a warning, as WebAssembly does not support "protected" visibility.

Changes to the Windows Target

• Produce COFF weak external symbols for IR level weak symbols without a comdat (e.g. for __attribute__((weak)) in C)

Changes to the DAG infrastructure

• A SelDag-level freeze instruction has landed. It is simply lowered as a copy operation to MachineIR, but to make it fully correct either IMPLICIT_DEF should be fixed or the equivalent FREEZE operation should be added to MachineIR.

Changes to the Debug Info

• LLVM now supports the debug entry values (DW_OP_entry_value) production for the x86, ARM, and AArch64 targets by default. Other targets can use the utility by using the experimental option ("-debug-entry-values"). This is a debug info feature that allows debuggers to recover the value of optimized-out parameters by going up a stack frame and interpreting the values passed to the callee. The feature improves the debugging user experience when debugging optimized code.

Changes to the Gold Plugin

• --plugin-opt=whole-program-visibility is added to specify that classes have hidden LTO visibility in LTO and ThinLTO links of source files compiled with -fwhole-program-vtables. See LTOVisibility for details. (D71913)

Changes to the LLVM tools

• Added an option (--show-section-sizes) to llvm-dwarfdump to show the sizes of all debug sections within a file.
• llvm-nm now implements the flag --special-syms and will filter out special symbols, i.e. mapping symbols on ARM and AArch64, by default. This matches the GNU nm behavior.
• llvm-rc now tolerates -1 as menu item ID, supports the language id option and allows string table values to be split into multiple string literals
• llvm-lib supports adding import library objects in addition to regular object files

External Open Source Projects Using LLVM 11

Zig Programming Language

Zig is a general-purpose programming language and toolchain for maintaining robust, optimal, and reusable software. In addition to supporting LLVM as an optional backend, Zig links Clang and LLD to provide an out-of-the-box cross compilation experience, not only for Zig code but for C and C++ code as well. Using a sophisticated caching system, Zig lazily builds from source compiler-rt, mingw-w64, musl, glibc, libcxx, libcxxabi, and libunwind for the selected target - a "batteries included" drop-in for GCC/Clang that works the same on every platform.

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 mailing lists.