LLVM 20.0.0git Release Notes ¶

Warning

These are in-progress notes for the upcoming LLVM 20.0.0git release. Release notes for previous releases can be found on the Download Page.

Introduction ¶

This document contains the release notes for the LLVM Compiler Infrastructure, release 20.0.0git. 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 ¶

Added a new IRNormalizer pass which aims to transform LLVM modules into a normal form by reordering and renaming instructions while preserving the same semantics. The normalizer makes it easier to spot semantic differences when diffing two modules which have undergone different passes.
…

Changes to the LLVM IR ¶

Types are no longer allowed to be recursive.
The x86_mmx IR type has been removed. It will be translated to the standard vector type <1 x i64> in bitcode upgrade.
Renamed llvm.experimental.stepvector intrinsic to llvm.stepvector.
Added usub_cond and usub_sat operations to atomicrmw.
Introduced noalias.addrspace metadata.
Remove the following intrinsics which can be replaced with a bitcast:
- llvm.nvvm.bitcast.f2i
- llvm.nvvm.bitcast.i2f
- llvm.nvvm.bitcast.d2ll
- llvm.nvvm.bitcast.ll2d
Remove the following intrinsics which can be replaced with a funnel-shift:
- llvm.nvvm.rotate.b32
- llvm.nvvm.rotate.right.b64
- llvm.nvvm.rotate.b64
Remove the following intrinsics which can be replaced with an addrspacecast:
- llvm.nvvm.ptr.gen.to.global
- llvm.nvvm.ptr.gen.to.shared
- llvm.nvvm.ptr.gen.to.constant
- llvm.nvvm.ptr.gen.to.local
- llvm.nvvm.ptr.global.to.gen
- llvm.nvvm.ptr.shared.to.gen
- llvm.nvvm.ptr.constant.to.gen
- llvm.nvvm.ptr.local.to.gen
Remove the following intrinsics which can be relaced with a load from addrspace(1) with an !invariant.load metadata
- llvm.nvvm.ldg.global.i
- llvm.nvvm.ldg.global.f
- llvm.nvvm.ldg.global.p
Operand bundle values can now be metadata strings.

Changes to LLVM infrastructure ¶

Changes to building LLVM ¶

Changes to TableGen ¶

Changes to Interprocedural Optimizations ¶

Changes to the AArch64 Backend ¶

.balign N, 0, .p2align N, 0, .align N, 0 in code sections will now fill the required alignment space with a sequence of 0x0 bytes (the requested fill value) rather than NOPs.
Assembler/disassembler support has been added for Armv9.6-A (2024) architecture extensions.

Changes to the AMDGPU Backend ¶

Removed llvm.amdgcn.flat.atomic.fadd and llvm.amdgcn.global.atomic.fadd intrinsics. Users should use the atomicrmw instruction with fadd and addrspace(0) or addrspace(1) instead.

Changes to the ARM Backend ¶

.balign N, 0, .p2align N, 0, .align N, 0 in code sections will now fill the required alignment space with a sequence of 0x0 bytes (the requested fill value) rather than NOPs.
The default behavior for frame pointers in leaf functions has been updated. When the -fno-omit-frame-pointer option is specified, FPKeepKindStr is set to -mframe-pointer=all, meaning the frame pointer (FP) is now retained in leaf functions by default. To eliminate the frame pointer in leaf functions, you must explicitly use the -momit-leaf-frame-pointer option.
When using the MOVT or MOVW instructions, the Assembler will now check to ensure that any addend that is used is within a 16-bit signed value range. If the addend falls outside of this range, the LLVM backend will emit an error like so Relocation Not In Range.

Changes to the AVR Backend ¶

Changes to the DirectX Backend ¶

Changes to the Hexagon Backend ¶

Changes to the LoongArch Backend ¶

Changes to the MIPS Backend ¶

Changes to the PowerPC Backend ¶

The Linux ppc64 LLC default cpu is updated from ppc to ppc64.
The AIX LLC default cpu is updated from generic to pwr7.

Changes to the RISC-V Backend ¶

.balign N, 0, .p2align N, 0, .align N, 0 in code sections will now fill the required alignment space with a sequence of 0x0 bytes (the requested fill value) rather than NOPs.
Added Syntacore SCR4 and SCR5 CPUs: -mcpu=syntacore-scr4/5-rv32/64
-mcpu=sifive-p470 was added.
Added Hazard3 CPU as taped out for RP2350: -mcpu=rp2350-hazard3 (32-bit only).
Fixed length vector support using RVV instructions now requires VLEN>=64. This means Zve32x and Zve32f will also require Zvl64b. The prior support was largely untested.
The Zvbc32e and Zvkgs extensions are now supported experimentally.
Added Smctr, Ssctr and Svvptc extensions.
-mcpu=syntacore-scr7 was added.
-mcpu=tt-ascalon-d8 was added.
The Zacas extension is no longer marked as experimental.
Added Smdbltrp, Ssdbltrp extensions to -march.
The Smmpm, Smnpm, Ssnpm, Supm, and Sspm pointer masking extensions are no longer marked as experimental.
The Sha extension is now supported.
The RVA23U64, RVA23S64, RVB23U64, and RVB23S64 profiles are no longer marked as experimental.
.insn <length>, <raw encoding> can be used to assemble 48- and 64-bit instructions from raw integer values.
.insn [<length>,] <raw encoding> now accepts absolute expressions for both expressions, so that they can be computed from constants and absolute symbols.
The following new inline assembly constraints and modifiers are accepted:
- cr constraint meaning an RVC-encoding compatible GPR (x8-x15)
- cf constraint meaning an RVC-encoding compatible FPR (f8-f15)
- R constraint meaning an even-odd GPR pair (prints as the even register, but both registers in the pair are considered live).
- N modifer meaning print the register encoding (0-31) rather than the name.
f and cf inline assembly constraints, when using F-/D-/H-in-X extensions, will use the relevant GPR rather than FPR. This makes inline assembly portable between e.g. F and Zfinx code.

Changes to the WebAssembly Backend ¶

The default target CPU, “generic”, now enables the -mnontrapping-fptoint and -mbulk-memory flags, which correspond to the Bulk Memory Operations and Non-trapping float-to-int Conversions language features, which are widely implemented in engines.

Changes to the Windows Target ¶

Changes to the X86 Backend ¶

.balign N, 0x90, .p2align N, 0x90, and .align N, 0x90 in code sections now fill the required alignment space with repeating 0x90 bytes, rather than using optimised NOP filling. Optimised NOP filling fills the space with NOP instructions of various widths, not just those that use the 0x90 byte encoding. To use optimised NOP filling in a code section, leave off the “fillval” argument, i.e. .balign N, .p2align N or .align N respectively.
Due to the removal of the x86_mmx IR type, functions with x86_mmx arguments or return values will use a different, incompatible, calling convention ABI. Such functions are not generally seen in the wild (Clang never generates them!), so this is not expected to result in real-world compatibility problems.
Support ISA of AVX10.2-256 and AVX10.2-512.
Supported instructions of MOVRS AND AVX10.2
Supported ISA of SM4(EVEX).
Supported ISA of MSR_IMM.
Supported -mcpu=diamondrapids

Changes to the OCaml bindings ¶

Changes to the Python bindings ¶

Changes to the C API ¶

The following symbols are deleted due to the removal of the x86_mmx IR type:
- LLVMX86_MMXTypeKind
- LLVMX86MMXTypeInContext
- LLVMX86MMXType
The following functions are added to further support non-null-terminated strings:
LLVMGetNamedFunctionWithLength
LLVMGetNamedGlobalWithLength
The following functions are added to access the LLVMContextRef associated with LLVMValueRef and LLVMBuilderRef objects:
- LLVMGetValueContext
- LLVMGetBuilderContext
The new pass manager can now be invoked with a custom alias analysis pipeline, using the LLVMPassBuilderOptionsSetAAPipeline function.
It is now also possible to run the new pass manager on a single function, by calling LLVMRunPassesOnFunction instead of LLVMRunPasses.
Support for creating instructions with custom synchronization scopes has been added:
- LLVMGetSyncScopeID to map a synchronization scope name to an ID.
- LLVMBuildFenceSyncScope, LLVMBuildAtomicRMWSyncScope and LLVMBuildAtomicCmpXchgSyncScope versions of the existing builder functions with an additional synchronization scope ID parameter.
- LLVMGetAtomicSyncScopeID and LLVMSetAtomicSyncScopeID to get and set the synchronization scope of any atomic instruction.
- LLVMIsAtomic to check if an instruction is atomic, for use with the above functions. Because of backwards compatibility, LLVMIsAtomicSingleThread and LLVMSetAtomicSingleThread continue to work with any instruction type.
The LLVMSetPersonalityFn and LLVMSetInitializer APIs now support clearing the personality function and initializer respectively by passing a null pointer.
The following functions are added to allow iterating over debug records attached to instructions:
- LLVMGetFirstDbgRecord
- LLVMGetLastDbgRecord
- LLVMGetNextDbgRecord
- LLVMGetPreviousDbgRecord
Added LLVMAtomicRMWBinOpUSubCond and LLVMAtomicRMWBinOpUSubSat to LLVMAtomicRMWBinOp enum for AtomicRMW instructions.

Changes to the CodeGen infrastructure ¶

Changes to the Metadata Info ¶

Changes to the Debug Info ¶

Changes to the LLVM tools ¶

Changes to LLDB ¶

LLDB now now supports inline diagnostics for the expression evaluator and command line parser.

Old:

(lldb) p a+b
error: <user expression 0>:1:1: use of undeclared identifier 'a'
    1 | a+b
      | ^
error: <user expression 0>:1:3: use of undeclared identifier 'b'
    1 | a+b
      |   ^

New:

(lldb) p a+b
         ˄ ˄
         │ ╰─ error: use of undeclared identifier 'b'
         ╰─ error: use of undeclared identifier 'a'

LLDB can now read the fpmr register from AArch64 Linux processes and core files.
Program stdout/stderr redirection will now open the file with O_TRUNC flag, make sure to truncate the file if path already exists.
- eg. settings set target.output-path/target.error-path <path/to/file>
A new setting target.launch-working-dir can be used to set a persistent cwd that is used by default by process launch and run.
LLDB now parses shared libraries in parallel, resulting in an average 2x speedup when attaching (only available on Darwin platforms) and launching (available on all platforms).

On the command line, LLDB now limits tab completions to your terminal width to avoid wrapping.