farmhash

Author: Ludovic HENRY ludovic.henry@qti.qualcomm.com
Date: 2026-08-14
Scope: RISC-V (riscv64/linux) support status for farmhash
Audience: Technical leadership, resource allocation strategy
Verification policy: Every claim is cross-referenced to a primary upstream source. Items that could not be verified against a second source are marked [NEEDS VERIFICATION].

1. Project Overview

FarmHash is a family of non-cryptographic hash functions developed at Google, designed as a successor to CityHash. The library provides 32-bit, 64-bit, and 128-bit hash functions with multiple variants tuned for different hardware capabilities. The canonical source is google/farmhash.

Governance: No formal governance model. The project lives under the google GitHub organization with no MAINTAINERS, OWNERS, or CODEOWNERS file. License is MIT. Community contact is farmhash-discuss@googlegroups.com.

Corporate maintainers: Google is the sole organizational owner. The primary historical maintainer is Geoff Pike (geoffpike, Google), who authored the bulk of commits and merged all pull requests through 2019. A June 2026 merge (SHA 9d99331 and 84fead3) was performed by a Google engineer (“eustas”) doing housekeeping merges of two stale PRs. No external corporate contributors with sustained involvement exist.

RISE membership: FarmHash is not listed on riseproject.dev. The project has no connection to the RISE Project. Google is a RISE Premier Member but has not directed any RISE resources toward FarmHash.

Community culture on new ports: The project has accepted community portability patches (Haiku, DragonFly BSD, ppc64le, s390x, MinGW - all merged via PR). However, maintainer cadence is extremely low. The last substantive algorithmic commit predates 2020. The June 2026 activity was only merging two PRs that had been open for years. A RISC-V patch would likely be accepted if submitted, but response times are measured in years and there is no proactive effort.

Project activity: 38 total commits, 19 merged PRs, 22 open issues as of the report date. The project is effectively in maintenance mode.

2. Port History and Upstreaming Timeline

Date Event Source
2016 PR #7/#8 merged: fixed configure: error: cannot guess build type on aarch64 by updating config.guess/config.sub google/farmhash PR #7
2017 Issue #12 filed: Debian package FTBFS on s390x and mips; resolved in later packaging google/farmhash issue #12
2019 Last substantive algorithmic commit google/farmhash commits
2019-05-13 Debian snapshot date for current packaged version (0~git20190513.0d859a8) Debian tracker: farmhash
2026-06-05 Most recent commit: minor housekeeping merge of two stale PRs google/farmhash commits
Never RISC-V port initiated Confirmed: zero riscv/riscv64 commits, issues, or PRs in repository history

Key contributors: Geoff Pike (Google) for all core algorithmic work. Community contributors for platform portability (aarch64, BSD variants, MinGW).

Upstreaming status: There is nothing to upstream. No RISC-V work exists anywhere - not in forks, not in patches, not in issue discussions.

3. Upstream Support Tier

FarmHash has no formal platform tier policy. There is no PLATFORMS.md, SUPPORT.md, or equivalent document.

Evidence by support indicator:

Indicator amd64 arm64 riscv64
CI coverage None None None
Official binary releases None (no releases published) None None
SIMD-accelerated code path Yes (SSE4.2, AES-NI, AVX) No No
Autoconf triplet recognition Yes Yes Yes (config.sub/config.guess only)
Known build failures None None None
Distro packaging Yes Yes Yes (Debian/Ubuntu)
Release-blocking test suite No CI; no release process No CI No CI

The practical support tier for riscv64 is: compiles and runs via portable scalar fallback; no upstream testing; no upstream acknowledgment of the architecture. This is identical to arm64’s upstream status - distro-maintained, not upstream-maintained.

4. Technical Architecture and RISC-V-Specific Subsystems

FarmHash’s high-performance paths are exclusively x86 SIMD. The architecture-specific code is concentrated in dev/platform.cc and the variant implementation files.

Architecture dispatch mechanism: platform.cc defines runtime detection macros: x86_64, x86, can_use_ssse3, can_use_sse41, can_use_sse42, can_use_aesni, can_use_avx. On riscv64, all of these evaluate to 0 at compile time. No __riscv, __aarch64__, NEON, or RVV macros appear anywhere in the codebase.

Component-level status:

Component Purpose amd64 arm64 riscv64
farmhashna 64-bit portable scalar C++ scalar C++ scalar C++
farmhashxo 64-bit portable scalar C++ scalar C++ scalar C++
farmhashuo 64-bit portable scalar C++ scalar C++ scalar C++
farmhashmk 32-bit portable scalar C++ scalar C++ scalar C++
farmhashcc 32-bit CityHash-compatible scalar C++ scalar C++ scalar C++
farmhashsa 32-bit SSE4.2 SIMD (full) missing missing
farmhashsu 32-bit SSE4.2 + AES-NI SIMD (full) missing missing
farmhashte 64-bit SSE4.1 SIMD (full) missing missing
farmhashns 64-bit SSE4.2 SIMD (partial) missing missing
farmhashnt 32-bit SSE4.2 SIMD (partial) missing missing

Critical detail on “missing” components: The SSE4.2-gated files (farmhashsa.cc, farmhashsu.cc, farmhashte.cc) use a guard pattern where #if !can_use_sse42 (or !can_use_sse41) branches to FARMHASH_DIE_IF_MISCONFIGURED and returns a stub value. This is a compile-time death sentinel, not a graceful fallback. These files together contain 113 _mm_* intrinsic calls and __m128i references. The amalgamated src/farmhash.cc contains 166 x86 SIMD references and zero RISC-V, RVV, NEON, or aarch64 references.

No RVV, Zbb, Zbc, or Zba usage: There is no RISC-V Vector extension path, no Zbc (carry-less multiply, which would be relevant for CRC-style hash mixing), and no Zbb bitmanipulation usage. The performance gap versus x86 SSE4.2 is structural.

Endianness: riscv64 is little-endian. FarmHash’s endianness detection uses __BYTE_ORDER__/__ORDER_LITTLE_ENDIAN__ macros and falls back to #include <endian.h> on Linux. The source comment states “FarmHash was developed with little-endian architectures in mind.” No manual override is needed for riscv64.

5. Build System, Cross-Compilation, and Toolchain

Build system: GNU Autoconf/Automake exclusively. No CMake. The repository contains a pre-generated configure script. Build files: configure, configure.ac, Makefile.am, src/Makefile.am, m4/ macros. No Dockerfiles, no .devcontainer, no container build infrastructure of any kind.

Native build on riscv64:

./configure CXXFLAGS="-g -O2"
make all check

The x86-specific flags (-mavx, -maes, -msse4.2) must be omitted entirely - they are x86-only and cause compilation errors on riscv64.

Cross-compilation from x86_64 host:

./configure \
  --host=riscv64-linux-gnu \
  CXX=riscv64-linux-gnu-g++ \
  CXXFLAGS="-g -O2"
make all

The configure script supports --host= for cross-compilation via standard autoconf cross_compiling=yes path. No toolchain files needed; no CMake toolchain.

Running the test binary under QEMU:

qemu-riscv64 -L /usr/riscv64-linux-gnu src/farmhash_unittest

Required toolchain versions: configure.ac states AC_PREREQ([2.65]) and LT_PREREQ([2.2]). The farmhash.cc source uses __builtin_bswap32/__builtin_bswap64 under a guard of (defined(__GNUC__) && ((__GNUC__ == 4 && __GNUC_MINOR__ >= 8) || __GNUC__ >= 5)), requiring GCC >= 4.8 or any Clang. Any GCC in a modern riscv64-linux-gnu-g++ toolchain (GCC 10+) is sufficient.

ISA feature flags for riscv64: None of the x86 FARMHASH_ASSUME_* defines should be set. Specifically:

Define riscv64 guidance
-DFARMHASH_ASSUME_SSSE3=1 Do not set (x86 only)
-DFARMHASH_ASSUME_SSE41=1 Do not set
-DFARMHASH_ASSUME_SSE42=1 Do not set
-DFARMHASH_ASSUME_AESNI=1 Do not set
-DFARMHASH_ASSUME_AVX=1 Do not set
-DFARMHASH_BIG_ENDIAN=1 Do not set (riscv64 is little-endian)

Known build failures: None reported on riscv64. The Debian buildd (rv-osuosl-04) successfully built 0~git20190513.0d859a8-4+b2 with no failures recorded.

6. Feature Coverage and Gap Analysis vs arm64 and amd64

Functional gaps: None. All hash functions that compile on riscv64 produce correct output. The portable C++ fallback paths (farmhashna, farmhashxo, farmhashmk, farmhashcc, farmhashuo) are fully functional on riscv64. The SSE4.2-gated variants are absent on both riscv64 and arm64 - this is not a riscv64-specific regression.

Correctness caveat: Issue #37 (open, filed 2020-02-22) documents that FarmHash’s non-fingerprint Hash* functions are not cross-platform stable - results differ between platforms by design. Only Fingerprint* functions are guaranteed portable across architectures. This affects any user relying on hash consistency between x86 and riscv64 deployments and is not riscv64-specific. [NEEDS VERIFICATION: whether this affects all non-fingerprint variants equally or only specific ones.]

Performance gaps:

Metric amd64 (x86 SSE4.2) arm64 riscv64
FarmHash32 bulk speed 22,136 MiB/sec Data not available Data not available: no riscv64 benchmarks found in any public suite
FarmHash64 bulk speed 12,929 MiB/sec Data not available Data not available
FarmHash128 bulk speed 14,454 MiB/sec Data not available Data not available
SIMD acceleration SSE4.2, AES-NI, AVX None None
RVV acceleration N/A N/A None (not implemented)

Note: The SMHasher maintainer explicitly flags FarmHash as “not portable, too machine specific: 64 vs 32bit, old gcc” and marks the x86 figures as “machine-specific (x64 SSE4/AVX)”. The smhasher benchmark suite itself fails to build on riscv64 due to x86 inline assembly (open issue rurban/smhasher#317, opened 2025-04-13), so no riscv64 baseline measurement is possible from that tool without upstream fixes.

Security hardening gaps: Data not available: no analysis of security hardening flags (stack canaries, CFI, shadow stack) across architectures was found in the research.

Missing SIMD opportunities on riscv64: The RVV (RISC-V Vector) extension and the Zbc (carry-less multiply) extension in the B-extension suite could accelerate hash mixing. Zbc in particular maps directly to CRC-style operations that the SSE4.2 _mm_crc32_u64 intrinsic uses in farmhashsu. Neither has been exploited. There is no tracking issue acknowledging this gap.

7. CI/CD Infrastructure

The google/farmhash repository has no CI configuration of any kind. The root directory contains only autoconf/automake build files and source code. There is no .github directory (confirmed via GitHub API returning HTTP 404), no GitHub Actions workflows, and no configuration for Travis CI, CircleCI, Cirrus CI, AppVeyor, Jenkins, or any other CI provider.

CI attribute amd64 arm64 riscv64
CI provider None None None
Automated build No No No
Automated test No No No
RISE/Scaleway runner No No No
Hardware board No No No

The config.sub and config.guess files in the repository contain riscv32 | riscv64 strings - these are GNU build system triplet recognizers and are entirely unrelated to CI.

The only validated riscv64 builds come from the Debian and Ubuntu package build infrastructure, which is external to the upstream project.

8. Distribution and Release Status

Upstream releases: None. The repository has never published a tagged release with binary assets. The GitHub releases page is empty.

Debian: Package farmhash (version 0~git20190513.0d859a8-4+b2) is present in Debian sid with riscv64 status “Installed”, built on rv-osuosl-04. All 18 tracked architectures show “Installed”. Source: Debian tracker: farmhash.

Ubuntu 24.04 (Noble): Both libfarmhash0 and libfarmhash-dev are available in the universe repository for: amd64, arm64, armhf, ppc64el, riscv64, s390x. Source: Ubuntu Packages: farmhash.

PyPI: The package named “farmhash” does not exist on PyPI (HTTP 404). The related package “pyfarmhash” publishes only win_amd64 and manylinux_x86_64 wheels - no riscv64 wheels.

RISE wheel builder: Not listed. The RISE wheel builder at riseproject.gitlab.io does not include farmhash.

Arch Linux: Not in official Arch Linux or Arch Linux RISC-V repositories (archriscv.felixc.at returns no results).

Summary for a user who needs a working riscv64 binary: Install libfarmhash-dev from Debian sid or Ubuntu 24.04 universe. No other binary channel exists. The installed library uses the portable scalar C++ fallback - there is no SIMD acceleration.

9. Dependencies

FarmHash has no external library dependencies. The configure.ac contains no PKG_CHECK_MODULES or AC_CHECK_LIB calls beyond the C++ standard library.

Dependency Role riscv64 Build riscv64 Test riscv64 Release Blocking Issues
C++ standard library (stdint.h, stdlib.h, string.h, utility) Core types, memcpy, std::swap OK - fully portable OK N/A (header-only use) None
byteswap / endian headers (byteswap.h, endian.h, platform variants) Byte-swap primitives OK - GCC __builtin_bswap64 covers riscv64 before any platform header OK N/A None
x86 SIMD intrinsics (immintrin.h, nmmintrin.h, wmmintrin.h) SSE4.1/SSE4.2/AES-NI/AVX fast paths Not applicable - guard macros never defined on riscv64 Not applicable N/A None (falls back to portable C++ automatically)
GCC/Clang compiler __builtin_expect, __builtin_bswap32/64, __builtin_unreachable OK - all builtins supported on riscv64 in GCC 12+, Clang 16+ OK N/A None
autoconf / automake / libtool (build-time only) Configure and Makefile generation OK N/A N/A None

No JIT, crypto library, numerics library, or other dependency with architecture-specific behavior is present. Dependency analysis terminates at one level.

11. Known Bugs and Active Issues

ID Title Status Severity Notes
#37 FreeBSD produces wrong hashes Open (2020-02-22) Medium - correctness trap Non-fingerprint Hash* functions are not cross-platform stable by design; Fingerprint* functions are portable. Affects all non-x86 platforms including riscv64 for users expecting hash consistency.
#35 Crash in farmhash used by ml vision Open (2019-07-07) Low - no architecture attribution No riscv64 attribution; origin unclear.
#21 s390x build failure Status unknown Low Historical big-endian build failure; not relevant to riscv64 (little-endian).

RISC-V specific issues: Zero. All 22 issues (all states) and all 19 PRs were reviewed. No riscv or riscv64 mention exists in the issue tracker.

Correctness issues for riscv64: Issue #37 is the only open correctness concern. Users must use Fingerprint* functions, not Hash* functions, to obtain consistent results across x86 and riscv64 deployments.

12. Objections and Upstream Blockers

Stated objections: None on record. No issue or PR has raised RISC-V and been rejected.

Technical blockers: None for functional correctness. The portable C++ fallback compiles and runs correctly on riscv64 without modification. The only technical gap is the absence of RVV or Zbc acceleration, which is a performance matter, not a correctness blocker.

Organizational blockers: The project is effectively dormant. With 38 total commits and the last algorithmic work dating to 2019, the maintainer bandwidth for reviewing and merging new architecture work is minimal. A submitted RISC-V patch would wait months to years for review based on historical PR merge latency (PRs #25 and #39 were merged in 2026 after being open for approximately 5-7 years). [NEEDS VERIFICATION: exact open dates for PRs #25 and #39.]

Acceptance probability: High for a well-formed patch that follows the existing x86 guard pattern and does not touch portable paths. The project has accepted community architecture ports (aarch64, s390x, Haiku, BSD variants). The barrier is maintainer response time, not technical opposition.

13. Investment Analysis

RISE has done no work on farmhash. The RISE wheel builder does not include it. No RISE blog post, funded effort, or tracked issue references farmhash. All investment sizing below represents net new work.

13.1 Functional Enablement

No work required. FarmHash builds and runs correctly on riscv64 via the portable scalar C++ fallback today. Debian and Ubuntu ship working riscv64 packages. There is no functional gap to close.

13.2 Performance Optimization

The performance gap is real but the project context limits the return on investment. The high-performance SSE4.2 paths that exist for x86 have no RISC-V equivalents. Potential optimizations:

  1. RVV-accelerated hash mixing: The SSE4.2 variants use 128-bit SIMD for parallel multiply-add operations. RVV equivalents are possible. Effort: 3-5 person-weeks for implementation, 2-3 person-weeks for validation against SMHasher quality tests.

  2. Zbc (carry-less multiply) for CRC-style mixing: The farmhashsu variant uses _mm_crc32_u64 (SSE4.2 CRC32). The Zbc extension provides clmul/clmulh instructions that could serve as an analog. Effort: 1-2 person-weeks.

Caveat: Given the project’s dormant state and the 5+ year PR merge latency, any optimization work faces an upstream absorption problem. Work may sit unmerged indefinitely unless a downstream fork or distro patch is acceptable as the deployment mechanism.

13.3 CI/CD Infrastructure

The project has no CI for any architecture. Adding riscv64 CI to a project that has no CI at all requires first establishing CI infrastructure, which is unlikely to be accepted upstream given the maintainer’s demonstrated non-engagement with CI tooling. Effort: 1-2 person-weeks to write GitHub Actions YAML; absorption probability low.

13.4 Ecosystem Enablement

FarmHash is a C++ library with no Python package (PyPI 404), no npm package, no Maven JAR, and no independent package ecosystem requiring enablement. It is consumed as a system library or as a vendored header. No ecosystem work is applicable.

13.5 Summary Table

Area Work Item Effort (person-weeks) Owner Priority
Functional None required - scalar fallback is fully functional 0 N/A N/A
Performance RVV-accelerated hash variants (farmhashsa/su/te equivalents) 5-8 Qualcomm / RISE contributor Low
Performance Zbc carry-less multiply for CRC-style mixing in farmhashsu equivalent 1-2 Qualcomm / RISE contributor Low
CI/CD Add riscv64 GitHub Actions job 1-2 RISE / contributor Low
Ecosystem N/A (no dependent package ecosystem) 0 N/A N/A

Overall investment recommendation: Defer. FarmHash works on riscv64 today with no blocking issues. The project is dormant and upstream absorption of optimization work is uncertain given multi-year PR merge latency. Downstream consumers (TensorFlow Lite / LiteRT notes farmhash as an internal dependency) obtain riscv64 support via the portable path with no issues. Invest in FarmHash only if a specific performance-sensitive workload on riscv64 hardware demonstrates a measurable hash throughput bottleneck.

14. Updates

No updates yet - initial report dated 2026-08-14.

15. References