4.3. Changelog
Note
To check the known issues encountered after the toolchain release, please refer to the content at the following link:
GNU Known Issues: Known Issues
LLVM Known Issues: Known Issues
libncrt Known Issues: Known Issues
4.3.1. Version 2026.10
Note
This version is modified based on version 2025.10; all tool components kept the same versions
Merged important upstream bugfixes, such as the VXRM state restore fix, the constant synthesis codegen fixes, and the Newlib
sqrt/sqrtfFPU fixAdded support for new extensions, such as
Zibi(v0.1),RNMI(Smrnmiv1.0),Double Trap(Smdbltrp/Ssdbltrpv1.0), and enhanced the Nuclei customXxlczextension support and theXxldspauto-generation capability
Toolchain components are unchanged from 2025.10: GCC 14.2.1, Binutils 2.44, GDB 16.2, Newlib 4.5.0, LLVM 19.1.7, and GLIBC 2.41.
Feature: Added
RNMI & Double Trapextension support (Smrnmiv1.0,Smdbltrp/Ssdbltrpv1.0) in GCC/LLVM/Binutils/GDB.Feature: Added
Zibiextension support (v0.1) in GCC/Binutils/GDB.Feature: Added new option
-mautovec-rvv/-mno-autovec-rvvto enable (default) or disable RVV auto-vectorization.Feature: The Nuclei DSP auto-vectorization (
-mautovec-dsp) and the NucleiXxlczauto-generation (-mautogen-xxlcz) are disabled by default now, and Binutils adds the same-mautogen-xxlcz/-mno-autogen-xxlczassembler option. The-maddibne/-mno-addibneoption is removed, thexl.addibneauto-generation is now controlled by-mautogen-xxlcz.Feature: Added GDB Python support, which can be enabled with
--with-pythonwhen building the toolchain.Feature: GDB now supports reading memory variables while the target is running, which allows Nuclei Studio 2026.10 and later versions to view variables without interrupting the target.
Feature: Updated the Nuclei custom ECLIC V2 CSRs in Binutils/GDB/LLVM.
Feature: GCC added the new constraint
Rfor register even-odd pairs.Feature: Interrupt handlers are RVV-free now; no RVV instructions will be generated in them.
Feature: Improved the implementation of the Nuclei
Xxlvqmaccextension, and added thevsetvliinstructions required by theXxlvwextension intrinsic functions.Feature: Enhanced the Nuclei
XxldspandXxlczextension capabilities.Feature: Updated
libncrtwith RISC-V big-endian andZicfilplanding pad support, see libncrt changelog Version 2026.10 for details.Fixed: The link failure when RVV intrinsics are used in LTO builds (PR target/110812).
Fixed: VXRM state not restored after calls (PR target/123626).
Fixed:
get_biv_step_1ICE for RISC-V (PR target/117506).Fixed: The missing left shift in the RISC-V constant synthesis (PR target/117690), which was a regression introduced in the 2025.10 toolchain.
Fixed: Newlib
sqrt/sqrtfused FPU when theForDextension is present.Fixed: The
xl.addibnescale field error in Binutils/GDB.Fixed: The error when compiling pure assembly files with
Xxlczin-march=, which existed since the 2025.02 toolchain.Fixed: The null pointer issue in
riscv_elf_after_check_relocsin Binutils.Fixed: The abort when displaying data and partial instructions in Binutils.
Fixed: The
objdump -d -l/-Sout of memory issue in Binutils.Fixed: The GDB disassemble core dump issue when the section name is NULL.
Fixed: Zcmt-related issues in Binutils/GDB, such as adjusting the Zcmt relax pass, fixing the conditions for automatic generation of Zcmt, fixing the Zcmt GDB table jump index, and adding the
--relax-zcmt/--no-relax-zcmt(default) linker options to control the Zcmt table jump relaxation.Fixed: The GCC
nuclei-900-seriespipeline model was 1-issue (single-issue) before, and is now corrected to 2-issue (dual-issue).LLVM: Added inline assembly support for GPR Pairs (
R).LLVM: Added Nuclei
Xxldspintrinsic support andXxlvfbfvector bf16 intrinsic support.LLVM: Reworked the Nuclei
-mtunesupport, which fixes issue 37 and supports the Nuclei 100/200/300/600/900/1000 series.LLVM: Fixed the llvm-objdump error and the Nuclei
Xxldspdisassembler issues.LLVM: Fixed the issue that
objdump -dwrongly disassembles 32-bit instructions as 16-bit, by adding the RISC-V mapping symbol ($x/$d) support to llvm-objdump.LLVM: Fixed the wrong
_mcountbehavior of the gprof instrumentation (-pg): theselfpc(return addressra) passed to_mcountwas incorrect, which made the callgraph unable to be generated correctly (upstream issue 121103).
4.3.2. Version 2025.10
Note
This version is modified based on version 2025.02; only Newlib and GLIBC received upstream version bumps
Windows toolchain now changed from Win32 to Win64
A lot of known issues are fixed in this release
Updated toolchain components to these versions: GCC 14.2.1, Binutils 2.44, GDB 16.2, Newlib 4.5.0, LLVM 19.1.7, and GLIBC 2.41.
Feature: Windows Newlib GCC now supports Win32_x64 (64-bit), replacing Win32 (32-bit).
Feature: Added auto-generation support for the
Zcmtextension.Feature: Initial support for
ECLIC V2instructions and CSRs added to Binutils and GDB.Feature: Added multilib support for
Zfinx,Zdinx, and related extensions.Feature: Binutils and GDB now add support for
Virtual Supervisor-levelCSRs.Feature: Add missing
fcvt.d.h/fcvt.h.dinstructions forXxlfbfextension.Feature: Added
Xxlvwextension support.Feature: Add a new
-maddibne/-mno-addibneoption to control whether Xxlcz’sxl.addibneinstruction is auto-generated.Performance optimization: Aligned Newlib memory/string routines (
memcpy/memset/setjmp/strcmp/strcpy/strlen) to 4-byte boundaries.Fixed: Compilation issues with semihosting functions
_timesand_gettimeofdayin Newlib.Fixed: Incorrect inline assembly translation for the
clrovinstruction inXxldsp(Binutils).Fixed: Conflicts between the
Xxlvfbfextension andZvfbfmin/Zvfbfwmaextensions.Fixed: Codegen regression in branch condition optimization, fixed by removing a redundant AND -1 instruction.
Fixed: Unnecessary
slli/sraiinstruction generation when using int16 type.LLVM: Initial implementation of pipeline support for Nuclei processor
100/200/300/600/900/1000series.LLVM: Added scalar BFloat16 support for Nuclei processors.
LLVM: Implemented Nuclei-specific VPU matrix instructions.
LLVM: Added missing
Virtual Supervisor-levelCSRs (hedelegh/medelegh) support.LLVM: Added Nuclei custom macro
__riscv_dsp, which can be used instead of__riscv_xxldspin the old version.LLVM: Added missing instructions
dkabs32,dsma32.u,clrov,rdov,kmada32, andsmbb32.LLVM: Fixed the issue where the
Xxldspextension could not be disassembled correctly.
4.3.3. Version 2025.02
Updated toolchain components to these versions: GCC 14.2.1, Binutils 2.44, GDB 16.2, Newlib 4.4.0, LLVM 19.1.7, and GLIBC 2.40.
Support for the
ZilsdandZclsdextensions.Some Nuclei custom CSR naming has been re-revised and corrected.
Implemented custom VPU intrinsics for Nuclei RISC-V CPU.
Nuclei introduced the
bf16type and support for theXxlvfbfandXxlfbfextensions.Added support for 3-issue (
nuclei-1000-3w-series) and 4-issue (nuclei-1000-4w-series) in the Nuclei 1000 series CPUs.GCC 14 introduces additional function attribute checks compared to GCC 13. For more details, you can refer to https://gcc.gnu.org/gcc-14/porting_to.html.
Add the option to automatically generate control for
Xxldspwith-mautovec-dsp/-mno-autovec-dspfor GCC, which is enabled by default.The
riscv_vector.hmust be included when using intrinsic type(s) and API(s). The scope of this attribute should not exceed the function body. Meanwhile, to make RVV types and API(s) available for this attribute, includingriscv_vector.hwill not report an error for now ifvis not present in-march=.LLVM: Add Zilsd & Zclsd v1.0 assembly support.
LLVM: Add Nuclei
Xxldsp/Xxldspn1x/Xxldspn2x/Xxldspn3xextensions assembly support.LLVM: Add Nuclei
Xxlczextension assembly support.LLVM: Add Nuclei
Xxlvqmaccextension intrinsic support.LLVM: Update Nuclei custom CSRs.
LLVM: Update the multilib list.
4.3.4. Version 2024.06
Updated toolchain components to these versions: GCC 13.1, Binutils 2.40, GDB 13.2, Newlib 4.3.0, LLVM 17.0.2, and GLIBC 2.38.
Instead of using single-letter
bkpto enable these extensions as we did on GCC 10, we split them all into corresponding sub-extensions, for example,_zba_zkr_zve32f, please check https://doc.nucleisys.com/nuclei_sdk/develop/buildsystem.html#arch-ext to learn about how to adapt Nuclei SDK to support GCC 13 upgraded from GCC 10.Implemented a new style of architecture extension test macros: each architecture extension has a corresponding feature test macro, which can be used to test its existence and version information. In addition, we add several custom macros,
__riscv_dsp,__riscv_bitmanip.Add new option
-misa-spec=*to control ISA spec version. This controls the default version of each extension. The official version is20191213, but it is set to2.2when configuring the Nuclei toolchain. The difference between them is that in the20191213version,ZicsrandZifenceiare separated from theiextension into two independent extensions, and using-misa-spec=2.2can avoid incompatible errors when theZicsrandZifenceiare not passed to-march=. For details, see https://github.com/riscv-collab/riscv-gnu-toolchain/issues/1315.Support for vector intrinsics as specified in version 0.12 of the RISC-V vector intrinsic specification.
The toolchain component prefix is
riscv-nuclei-elf-on GCC 10, but isriscv64-unknown-elf-on GCC 13.On GCC 10, RISC-V intrinsic API heads contain
riscv_vector.h,riscv_vector_itr.h,rvintrin.h,rvp_intrinsic.h, but now onlyriscv_vector.h,rvp_intrinsic.h,riscv_nuclei_xlcz.hare provided in GCC 13, if you want to findborkintrinsic API, please check https://github.com/riscv/riscv-crypto/blob/main/benchmarks/share/rvintrin.h and https://github.com/riscv/riscv-crypto/blob/main/benchmarks/share/riscv-crypto-intrinsics.h, and for RVV intrinsic API, we support 0.12 in GCC 13 now, see https://github.com/riscv-non-isa/rvv-intrinsic-doc/releases/tag/v1.0-rc0.The version of the libncrt was changed from v2.0.0 to v3.0.0, and libncrt is now split into three parts,
libncrt,heapopsandfileops, click https://doc.nucleisys.com/nuclei_sdk/develop/buildsystem.html#stdclib to learn about how the Newlib/libncrt are used in Nuclei SDK with GCC 13.