[PATCH v6 0/2] riscv: lib: add optimized memcmp() and extend KUnit tests

[PATCH v6 0/2] riscv: lib: add optimized memcmp() and extend KUnit tests

[Milan Tripkovic]

This v6 series provides an optimized RISC-V memcmp() implementation 
along with extended KUnit tests for functional verification and 
benchmarking

The memcmp() implementation now uses word-aligned loads for the generic 
path and leverages Zbb for fast sign resolution on supported hardware. 
A robust fallback to byte-by-byte comparison is implemented for non-Zbb 
systems to ensure correctness and stability across different toolchains 
and hardware revisions.

Signed-off-by: Milan Tripkovic <Milan.Tripkovic@rt-rk.com>

---
v6 changes:
    - Fixed build errors kernel test robot noticed. 
    - Link to v5: https://lore.kernel.org/all/20260519140029.1190381-1-milant2002@gmail.com/
    
v5 changes:
    - Removed redundant warm-up from test case
    - Optimized generic memcmp path with word-aligned loads.
    - Implemented non-ZBB fallback to byte comparison per review.
    - Link to v4: https://lore.kernel.org/all/20260518131407.1026049-2-milant2002@gmail.com/

v4 changes:
    - Fixed build error: call to undeclared function 'STRING_BENCH'
    - Link to v3: https://lore.kernel.org/all/20260515141019.1234037-1-milant2002@gmail.com/

v3 changes:
    - Split memcmp benchmark into wrapper (string_bench_memcmp) and worker
      function (do_string_bench_memcmp).
    - Removed all C99 mixed declarations; moved all variable declarations
      to the top of each function.
    - Converted len, iterations and loop counters in the benchmark to u64
      to avoid implicit casts.
    - Cleaned up spacing, indentation and minor style issues.
    - Added #if defined(CONFIG_RISCV_ISA_ZBB)... in  memcmp.S
    - Link to v2: https://lore.kernel.org/all/20260514121359.931999-1-milant2002@gmail.com/

v2 changes:
    - Added alignment checks for buffers to avoid expensive misaligned loads.
    - Optimized the loop using end-pointers to reduce per-iteration overhead.
    - Implemented word-aligned tail handling using ZBB shifts.
    - Removed redundant pointer equality (a0 == a1) check.
    - Retained BE support via #ifndef; ZBB rev8 is used for the LE fast-path.
    - Fixed KUnit build failures for Clang and non-benchmark configs.
    - Link to v1: https://lore.kernel.org/all/20260512141007.1193033-1-milant2002@gmail.com/

Milan Tripkovic (2):
  riscv: lib: add memcmp() implementation
  lib/string_kunit: extend benchmarks and unit test to memcmp()

 arch/riscv/include/asm/string.h |   2 +
 arch/riscv/lib/Makefile         |   1 +
 arch/riscv/lib/memcmp.S         | 107 ++++++++++++++++++++++++++++++++
 arch/riscv/purgatory/Makefile   |   5 +-
 lib/tests/string_kunit.c        | 106 +++++++++++++++++++++++++++++++
 5 files changed, 220 insertions(+), 1 deletion(-)
 create mode 100644 arch/riscv/lib/memcmp.S

-- 
2.43.0

[PATCH v6 1/2] riscv: lib: add memcmp() implementation

[Milan Tripkovic]

From: Milan Tripkovic <Milan.Tripkovic@rt-rk.com>

Add an assembly implementation of memcmp() for RISC-V. The implementation
uses word-aligned loads for the generic path and the ZBB extension for
optimized sign resolution, with an assembly fallback for non-ZBB systems.

Benchmark results (QEMU TCG, rv64):

 Len   | Default |NoALIGN | ALIGN  | %NoALIGN | %ALIGN
 ------|---------|--------|--------|----------|--------
 1 B   | 20.3    | 21.5   | 20.9   | +5.9%    | +3.0%
 7 B   | 88.9    | 96.9   | 155.7  | +9.0%    | +75.1%
 8 B   | 89.6    | 110.5  | 176.2  | +23.3%   | +96.7%
 16 B  | 134.4   | 172.8  | 334.8  | +28.6%   | +149.1%
 31 B  | 163.5   | 211.9  | 606.2  | +29.6%   | +270.8%
 64 B  | 203.8   | 240.3  | 968.6  | +17.9%   | +375.3%
 127 B | 224.6   | 269.0  | 1362.8 | +19.8%   | +506.8%
 512 B | 235.7   | 269.9  | 1913.7 | +14.5%   | +711.9%
 1024 B| 256.8   | 283.5  | 2123.6 | +10.4%   | +726.9%
 4096 B| 263.8   | 299.7  | 2290.4 | +13.6%   | +768.2%

Signed-off-by: Milan Tripkovic <Milan.Tripkovic@rt-rk.com>
---
v6 changes:
    - Fixed build errors kernel test robot noticed. 
    - Link to v5: https://lore.kernel.org/all/20260519140029.1190381-1-milant2002@gmail.com/
    
v5 changes:
    - Optimized generic memcmp path with word-aligned loads.
    - Implemented non-ZBB fallback to byte comparison per review.
    - Link to v4: https://lore.kernel.org/all/20260518131407.1026049-2-milant2002@gmail.com/

v2 changes:
    - Added alignment checks for buffers to avoid expensive misaligned loads.
    - Optimized the loop using end-pointers to reduce per-iteration overhead.
    - Implemented word-aligned tail handling using ZBB shifts.
    - Removed redundant pointer equality (a0 == a1) check.
    - Retained BE support via #ifndef; ZBB rev8 is used for the LE fast-path.
    - Link to v1: https://lore.kernel.org/all/20260512141007.1193033-1-milant2002@gmail.com/

 arch/riscv/include/asm/string.h |   2 +
 arch/riscv/lib/Makefile         |   1 +
 arch/riscv/lib/memcmp.S         | 107 ++++++++++++++++++++++++++++++++
 arch/riscv/purgatory/Makefile   |   5 +-
 4 files changed, 114 insertions(+), 1 deletion(-)
 create mode 100644 arch/riscv/lib/memcmp.S

diff --git a/arch/riscv/include/asm/string.h b/arch/riscv/include/asm/string.h
index 764ffe8f6..5c5299678 100644
--- a/arch/riscv/include/asm/string.h
+++ b/arch/riscv/include/asm/string.h
@@ -18,6 +18,8 @@ extern asmlinkage void *__memcpy(void *, const void *, size_t);
 #define __HAVE_ARCH_MEMMOVE
 extern asmlinkage void *memmove(void *, const void *, size_t);
 extern asmlinkage void *__memmove(void *, const void *, size_t);
+#define __HAVE_ARCH_MEMCMP
+extern asmlinkage int memcmp(const void *, const void *, size_t);
 
 #if !(defined(CONFIG_KASAN_GENERIC) || defined(CONFIG_KASAN_SW_TAGS))
 #define __HAVE_ARCH_STRCMP
diff --git a/arch/riscv/lib/Makefile b/arch/riscv/lib/Makefile
index 6f767b2a3..b529e1be1 100644
--- a/arch/riscv/lib/Makefile
+++ b/arch/riscv/lib/Makefile
@@ -3,6 +3,7 @@ lib-y			+= delay.o
 lib-y			+= memcpy.o
 lib-y			+= memset.o
 lib-y			+= memmove.o
+lib-y			+= memcmp.o
 ifeq ($(CONFIG_KASAN_GENERIC)$(CONFIG_KASAN_SW_TAGS),)
 lib-y			+= strcmp.o
 lib-y			+= strlen.o
diff --git a/arch/riscv/lib/memcmp.S b/arch/riscv/lib/memcmp.S
new file mode 100644
index 000000000..bdb92941d
--- /dev/null
+++ b/arch/riscv/lib/memcmp.S
@@ -0,0 +1,107 @@
+/* SPDX-License-Identifier: GPL-2.0-only */
+
+#include <linux/linkage.h>
+#include <asm/asm.h>
+#include <asm/alternative-macros.h>
+#include <asm/hwcap.h>
+
+#if defined(CONFIG_RISCV_ISA_ZBB) && defined(CONFIG_TOOLCHAIN_HAS_ZBB)
+.option push
+.option arch,+zbb
+#endif
+
+/* int memcmp(const void *cs, const void *ct, size_t n) */
+SYM_FUNC_START(memcmp)
+/*
+ * Parameters
+ *	a0 - Pointer to first memory block (cs), also return value
+ *	a1 - Pointer to second memory block (ct)
+ *	a2 - Number of bytes to compare (n), decremented during loop
+ *
+ * Returns
+ *	a0 - 0 if equal, positive if cs > ct, negative if cs < ct
+ *
+ * Clobbers
+ *	t0, t1, t2, t3, t4
+ */
+	add	t3, a0, a2
+	or	t0, a0, a1
+	andi	t0, t0, (SZREG - 1)
+	bnez	t0, 5f
+
+	addi	t4, t3, -SZREG
+	bltu	t4, a0, 7f
+
+1:
+	REG_L	t1, 0(a0)
+	REG_L	t2, 0(a1)
+	bne	t1, t2, 2f
+	addi	a0, a0, SZREG
+	addi	a1, a1, SZREG
+	bleu	a0, t4, 1b
+
+7:
+#if defined(CONFIG_RISCV_ISA_ZBB) && defined(CONFIG_TOOLCHAIN_HAS_ZBB)
+	__ALTERNATIVE_CFG("j 5f", "nop", 0, RISCV_ISA_EXT_ZBB,
+			  IS_ENABLED(CONFIG_RISCV_ISA_ZBB))
+
+	beq	a0, t3, 4f
+	REG_L	t1, 0(a0)
+	REG_L	t2, 0(a1)
+
+	sub	t0, t3, a0
+	li	t4, SZREG
+	sub	t0, t4, t0
+	slli	t0, t0, 3
+
+#ifndef CONFIG_CPU_BIG_ENDIAN
+	rev8	t1, t1
+	rev8	t2, t2
+#endif
+	srl	t1, t1, t0
+	srl	t2, t2, t0
+
+	bne	t1, t2, 8f
+	li	a0, 0
+	ret
+#else
+	j	5f
+#endif
+5:
+	beq	a0, t3, 4f
+6:
+	lbu	t1, 0(a0)
+	lbu	t2, 0(a1)
+	bne	t1, t2, 3f
+	addi	a0, a0, 1
+	addi	a1, a1, 1
+	bne	a0, t3, 6b
+
+4:	li	a0, 0
+	ret
+2:
+#if !defined(CONFIG_CPU_BIG_ENDIAN) && \
+	defined(CONFIG_TOOLCHAIN_HAS_ZBB) && defined(CONFIG_RISCV_ISA_ZBB)
+	__ALTERNATIVE_CFG("j 5b", "nop", 0, RISCV_ISA_EXT_ZBB, IS_ENABLED(CONFIG_RISCV_ISA_ZBB))
+	rev8	t1, t1
+	rev8	t2, t2
+#elif !defined(CONFIG_CPU_BIG_ENDIAN)
+	j	5b
+#endif
+8:
+	sltu	a0, t2, t1
+	sltu	t0, t1, t2
+	sub	a0, a0, t0
+	ret
+
+3:
+	sub	a0, t1, t2
+	ret
+
+SYM_FUNC_END(memcmp)
+SYM_FUNC_ALIAS(__pi_memcmp, memcmp)
+EXPORT_SYMBOL(memcmp)
+
+#if defined(CONFIG_RISCV_ISA_ZBB) && defined(CONFIG_TOOLCHAIN_HAS_ZBB)
+.option pop
+#endif
diff --git a/arch/riscv/purgatory/Makefile b/arch/riscv/purgatory/Makefile
index b0358a78f..456929971 100644
--- a/arch/riscv/purgatory/Makefile
+++ b/arch/riscv/purgatory/Makefile
@@ -1,6 +1,6 @@
 # SPDX-License-Identifier: GPL-2.0
 
-purgatory-y := purgatory.o sha256.o entry.o string.o ctype.o memcpy.o memset.o
+purgatory-y := purgatory.o sha256.o entry.o string.o ctype.o memcpy.o memset.o memcmp.o
 ifeq ($(CONFIG_KASAN_GENERIC)$(CONFIG_KASAN_SW_TAGS),)
 purgatory-y += strcmp.o strlen.o strncmp.o strnlen.o strchr.o strrchr.o
 endif
@@ -41,6 +41,9 @@ $(obj)/strchr.o: $(srctree)/arch/riscv/lib/strchr.S FORCE
 $(obj)/strrchr.o: $(srctree)/arch/riscv/lib/strrchr.S FORCE
 	$(call if_changed_rule,as_o_S)
 
+$(obj)/memcmp.o: $(srctree)/arch/riscv/lib/memcmp.S FORCE
+	$(call if_changed_rule,as_o_S)
+
 CFLAGS_sha256.o := -D__DISABLE_EXPORTS -D__NO_FORTIFY
 CFLAGS_string.o := -D__DISABLE_EXPORTS
 CFLAGS_ctype.o := -D__DISABLE_EXPORTS
-- 
2.43.0

[PATCH v6 2/2] lib/string_kunit: extend benchmarks and unit test to memcmp()

[Milan Tripkovic]

From: Milan Tripkovic <Milan.Tripkovic@rt-rk.com>

Extend the string benchmarking suite to include memcmp().
Extend the string unit test to include memcmp().

Signed-off-by: Milan Tripkovic <Milan.Tripkovic@rt-rk.com>
---
v6 changes:
    - Fixed build errors kernel test robot noticed. 
    - Link to v5: https://lore.kernel.org/all/20260519140029.1190381-1-milant2002@gmail.com/
    
v5 changes:
    - Removed redundant warm-up from test case
    - Link to v4: https://lore.kernel.org/all/20260518131407.1026049-2-milant2002@gmail.com/

v4 changes:
    - Fixed build error: call to undeclared function 'STRING_BENCH'
    - Link to v3: https://lore.kernel.org/all/20260515141019.1234037-1-milant2002@gmail.com/

v3 changes:
    - Split memcmp benchmark into wrapper (string_bench_memcmp) and worker
      function (do_string_bench_memcmp).
    - Removed all C99 mixed declarations; moved all variable declarations
      to the top of each function.
    - Converted len, iterations and loop counters in the benchmark to u64
      to avoid implicit casts.
    - Cleaned up spacing, indentation and minor style issues.
    - Added #if defined(CONFIG_RISCV_ISA_ZBB)... in  memcmp.S
    - Link to v2: https://lore.kernel.org/all/20260514121359.931999-1-milant2002@gmail.com/

v2 changes:
    - Fixed KUnit build failures for Clang and non-benchmark configs.
    - Link to v1: https://lore.kernel.org/all/20260512141007.1193033-1-milant2002@gmail.com/
    
 lib/tests/string_kunit.c | 106 +++++++++++++++++++++++++++++++++++++++
 1 file changed, 106 insertions(+)

diff --git a/lib/tests/string_kunit.c b/lib/tests/string_kunit.c
index 0819ace5b..d395627f7 100644
--- a/lib/tests/string_kunit.c
+++ b/lib/tests/string_kunit.c
@@ -881,6 +881,110 @@ static void string_bench_strrchr(struct kunit *test)
 	STRING_BENCH_BUF(test, buf, len, strrchr, buf, '\0');
 }
 
+static void string_test_memcmp(struct kunit *test)
+{
+	const unsigned int max_offset = 16;
+	const unsigned int max_len = 32;
+	const unsigned int buf_size = max_offset + max_len + 32;
+	u8 *buf1, *buf2;
+	unsigned int i, j, len, k;
+	int res;
+
+	buf1 = kunit_kzalloc(test, buf_size, GFP_KERNEL);
+	buf2 = kunit_kzalloc(test, buf_size, GFP_KERNEL);
+	KUNIT_ASSERT_NOT_ERR_OR_NULL(test, buf1);
+	KUNIT_ASSERT_NOT_ERR_OR_NULL(test, buf2);
+
+	for (i = 0; i < max_offset; i++) {
+		for (j = 0; j < max_offset; j++) {
+			for (len = 0; len <= max_len; len++) {
+				memset(buf1, 'A', buf_size);
+				memset(buf2, 'A', buf_size);
+				KUNIT_EXPECT_EQ_MSG(test, memcmp(buf1 + i, buf2 + j, len), 0,
+						    "Should be equal: i:%u j:%u len:%u", i, j, len);
+				for (k = 0; k < len; k++) {
+					memset(buf1, 'A', buf_size);
+					memset(buf2, 'A', buf_size);
+					buf2[j + k] = 'B';
+					res = memcmp(buf1 + i, buf2 + j, len);
+					KUNIT_EXPECT_NE_MSG(test, res, 0,
+							    "Should detect difference at k:%u (i:%u j:%u len:%u)",
+						k, i, j, len);
+					if (buf1[i + k] < buf2[j + k])
+						KUNIT_EXPECT_LT(test, res, 0);
+					else
+						KUNIT_EXPECT_GT(test, res, 0);
+				}
+			}
+		}
+	}
+}
+
+#ifndef STRING_BENCH
+#define STRING_BENCH(...) 0
+#endif
+
+static void do_string_bench_memcmp(struct kunit *test)
+{
+	char *buf1 = NULL;
+	char *buf2 = NULL;
+	const u64 lengths[] = { 1, 7, 8, 16, 32, 64, 128, 512, 1024, 4096};
+	const int offsets[] = { 0, 1, 3, 7};
+	const u64 max_len = 4096 + 64;
+	unsigned int o, i;
+	unsigned int off;
+	u64 len;
+	u64 iterations;
+	u64 elapsed;
+	u64 ns_per_call;
+	u64 mbps;
+
+	buf1 = vmalloc(max_len);
+	buf2 = vmalloc(max_len);
+
+	if (!buf1 || !buf2) {
+		vfree(buf1);
+		vfree(buf2);
+		kunit_err(test, "vmalloc failed\n");
+		return;
+	}
+
+	memset(buf1, 'A', max_len);
+	memset(buf2, 'A', max_len);
+
+	for (o = 0; o < ARRAY_SIZE(offsets); o++) {
+		off = offsets[o];
+
+		for (i = 0; i < ARRAY_SIZE(lengths); i++) {
+			len = lengths[i];
+			iterations = (len < 512) ? 100000ULL : 10000ULL;
+			elapsed = STRING_BENCH(iterations, memcmp, buf1, buf2 + off, len);
+			ns_per_call = div_u64(elapsed, iterations);
+			mbps = len ? div_u64(iterations * len * (NSEC_PER_SEC / MEGA), elapsed) : 0;
+
+			if (off == 0) {
+				kunit_info(test, "bench_memcmp_aligned: len=%-4llu: %llu MB/s (%llu ns/call)\n",
+					   len, mbps, ns_per_call);
+			} else {
+				kunit_info(test, "bench_memcmp_unaligned(off=%u): len=%-4llu: %llu MB/s (%llu ns/call)\n",
+					   off, len, mbps, ns_per_call);
+			}
+		}
+	}
+
+	vfree(buf1);
+	vfree(buf2);
+}
+
+static void string_bench_memcmp(struct kunit *test)
+{
+	if (!IS_ENABLED(CONFIG_STRING_KUNIT_BENCH)) {
+		kunit_skip(test, "CONFIG_STRING_KUNIT_BENCH not enabled");
+		return;
+	}
+	do_string_bench_memcmp(test);
+}
+
 static struct kunit_case string_test_cases[] = {
 	KUNIT_CASE(string_test_memset16),
 	KUNIT_CASE(string_test_memset32),
@@ -910,6 +1014,8 @@ static struct kunit_case string_test_cases[] = {
 	KUNIT_CASE(string_bench_strnlen),
 	KUNIT_CASE(string_bench_strchr),
 	KUNIT_CASE(string_bench_strrchr),
+	KUNIT_CASE(string_test_memcmp),
+	KUNIT_CASE_SLOW(string_bench_memcmp),
 	{}
 };
 
-- 
2.43.0

Re: [PATCH v6 2/2] lib/string_kunit: extend benchmarks and unit test to memcmp()

[Paul Walmsley]

On Wed, 20 May 2026, Milan Tripkovic wrote:

> From: Milan Tripkovic <Milan.Tripkovic@rt-rk.com>
> 
> Extend the string benchmarking suite to include memcmp().
> Extend the string unit test to include memcmp().
> 
> Signed-off-by: Milan Tripkovic <Milan.Tripkovic@rt-rk.com>

Same comments on this patch as were posted here:

  https://lore.kernel.org/linux-riscv/81c7c1b0-2906-96f6-6f7e-c859a7578c8c@kernel.org/T/#u


- Paul