This is an automated email from the ASF dual-hosted git repository.
chenBright pushed a commit to branch master
in repository https://gitbox.apache.org/repos/asf/brpc.git
The following commit(s) were added to refs/heads/master by this push:
new f16ab1c5 snappy: optimize UnalignedCopy64 and IncrementalCopy for
RISC-V (#3360)
f16ab1c5 is described below
commit f16ab1c581746919fb394c3cfd3735ed83a4e3bc
Author: Felix-Gong <[email protected]>
AuthorDate: Tue Jul 7 15:26:10 2026 +0800
snappy: optimize UnalignedCopy64 and IncrementalCopy for RISC-V (#3360)
* snappy: optimize UnalignedCopy64 and IncrementalCopy for RISC-V
Use RISC-V inline assembly (ld/sd) for 8-byte copy operations instead
of generic macro-based implementation.
Changes:
- UnalignedCopy64: direct ld/sd pair for 8-byte copy
- IncrementalCopy: 8-byte bulk copies when source/dest don't overlap
Performance improvement (direct function benchmark):
- Decompress compressible-256K: 728 MB/s -> 2205 MB/s (+203%)
- Decompress zeros-256K: 543 MB/s -> 1462 MB/s (+169%)
Tests: brpc_snappy_compress_unittest passed (7/7)
Signed-off-by: Xiaofei Gong <[email protected]>
Signed-off-by: YuanSheng <[email protected]>
* snappy: add alignment check for RISC-V ld/sd optimization
Address Copilot review: ld/sd require 8-byte alignment. Add runtime
alignment check and fall back to memcpy/byte-copy for unaligned
addresses to avoid traps on implementations that don't support
misaligned access.
Signed-off-by: Xiaofei Gong <[email protected]>
Signed-off-by: YuanSheng <[email protected]>
---------
Signed-off-by: Xiaofei Gong <[email protected]>
Signed-off-by: YuanSheng <[email protected]>
---
.../third_party/snappy/snappy-stubs-internal.h | 16 +++++++++++++++
src/butil/third_party/snappy/snappy.cc | 23 ++++++++++++++++++++++
2 files changed, 39 insertions(+)
diff --git a/src/butil/third_party/snappy/snappy-stubs-internal.h
b/src/butil/third_party/snappy/snappy-stubs-internal.h
index e94a9c73..764f97bc 100644
--- a/src/butil/third_party/snappy/snappy-stubs-internal.h
+++ b/src/butil/third_party/snappy/snappy-stubs-internal.h
@@ -164,6 +164,21 @@ inline void UNALIGNED_STORE64(void *p, uint64_t v) {
// This can be more efficient than UNALIGNED_LOAD64 + UNALIGNED_STORE64
// on some platforms, in particular ARM.
inline void UnalignedCopy64(const void *src, void *dst) {
+#if defined(__riscv) && __riscv_xlen == 64
+ // RISC-V optimized: single ld/sd pair for 8-byte copy (aligned only)
+ if ((((uintptr_t)src | (uintptr_t)dst) & 7) == 0) {
+ uint64_t tmp;
+ __asm__ volatile(
+ "ld %0, %1\n\t"
+ "sd %0, %2\n\t"
+ : "=&r"(tmp)
+ : "m"(*(const uint64_t*)src), "m"(*(uint64_t*)dst)
+ : "memory");
+ } else {
+ // Unaligned: fall back to memcpy-based approach
+ memcpy(dst, src, 8);
+ }
+#else
if (sizeof(void *) == 8) {
UNALIGNED_STORE64(dst, UNALIGNED_LOAD64(src));
} else {
@@ -173,6 +188,7 @@ inline void UnalignedCopy64(const void *src, void *dst) {
UNALIGNED_STORE32(dst_char, UNALIGNED_LOAD32(src_char));
UNALIGNED_STORE32(dst_char + 4, UNALIGNED_LOAD32(src_char + 4));
}
+#endif
}
// Convert to little-endian storage, opposite of network format.
diff --git a/src/butil/third_party/snappy/snappy.cc
b/src/butil/third_party/snappy/snappy.cc
index c42889f8..bcbf39ed 100644
--- a/src/butil/third_party/snappy/snappy.cc
+++ b/src/butil/third_party/snappy/snappy.cc
@@ -97,9 +97,32 @@ static const uint32_t kMaximumTagLength = 5; //
COPY_4_BYTE_OFFSET plus the act
// or memmove().
static inline void IncrementalCopy(const char* src, char* op, ssize_t len) {
assert(len > 0);
+#if defined(__riscv) && __riscv_xlen == 64
+ // RISC-V optimized: use 8-byte copies when aligned and safe
+ if (len >= 8 && (op - src >= 8 || src - op >= 8) &&
+ (((uintptr_t)src | (uintptr_t)op) & 7) == 0) {
+ do {
+ uint64_t tmp;
+ __asm__ volatile(
+ "ld %0, %1\n\t"
+ "sd %0, %2\n\t"
+ : "=&r"(tmp)
+ : "m"(*(const uint64_t*)src), "m"(*(uint64_t*)op)
+ : "memory");
+ src += 8;
+ op += 8;
+ len -= 8;
+ } while (len >= 8);
+ }
+ while (len > 0) {
+ *op++ = *src++;
+ --len;
+ }
+#else
do {
*op++ = *src++;
} while (--len > 0);
+#endif
}
// Equivalent to IncrementalCopy except that it can write up to ten extra
---------------------------------------------------------------------
To unsubscribe, e-mail: [email protected]
For additional commands, e-mail: [email protected]
