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

[PATCH 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 the ZBB extension for word-at-a-time comparison and an assembly
fallback for non-ZBB systems.

Benchmark results (QEMU TCG, rv64):

  Len  | Def   | NoZBB | ZBB   | %NoZBB | %ZBB
  -----|-------|-------|-------|--------|-------
  1 B  | 22.4  | 24.6  | 23.2  | +9.8%  | +3.5%
  7 B  | 96.9  | 108.5 | 107.3 | +12.0% | +10.7%
  8 B  | 107.0 | 116.3 | 176.7 | +8.7%  | +65.1%
  16 B | 148.4 | 172.8 | 315.6 | +16.4% | +112.6%
  31 B | 182.2 | 217.1 | 377.6 | +19.2% | +107.2%
  64 B | 220.6 | 239.4 | 874.2 | +8.5%  | +296.2%
  127 B| 213.7 | 254.8 | 1042.9| +19.2% | +388.0%
  512 B| 255.1 | 269.0 | 1778.6| +5.4%  | +597.2%
  1024B| 252.3 | 280.9 | 1887.7| +11.3% | +648.1%
  3173B| 241.3 | 288.7 | 2063.2| +19.6% | +755.0%
  4096B| 240.9 | 280.5 | 2064.5| +16.4% | +756.9%

Signed-off-by: Milan Tripkovic <Milan.Tripkovic@rt-rk.com>
---
 arch/riscv/include/asm/string.h |   2 +
 arch/riscv/lib/Makefile         |   1 +
 arch/riscv/lib/memcmp.S         | 103 ++++++++++++++++++++++++++++++++
 arch/riscv/purgatory/Makefile   |   5 +-
 4 files changed, 110 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..444b082d9
--- /dev/null
+++ b/arch/riscv/lib/memcmp.S
@@ -0,0 +1,103 @@
+/* SPDX-License-Identifier: GPL-2.0-only */
+
+#include <linux/linkage.h>
+#include <asm/asm.h>
+#include <asm/alternative-macros.h>
+#include <asm/hwcap.h>
+
+/* int memcmp(const void *cs, const void *ct, size_t n) */
+SYM_FUNC_START(memcmp)
+
+	__ALTERNATIVE_CFG("nop", "j memcmp_zbb", 0, RISCV_ISA_EXT_ZBB,
+		IS_ENABLED(CONFIG_RISCV_ISA_ZBB) && IS_ENABLED(CONFIG_TOOLCHAIN_HAS_ZBB))
+/*
+ * 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), transformed to end pointer (a0 + n)
+ *
+ * Returns
+ *	a0 - 0 if equal, positive if cs > ct, negative if cs < ct
+ *
+ * Clobbers
+ *	t0, t1
+ */
+	beqz	a2, 2f
+	add	a2, a0, a2
+1:
+	lbu	t0, 0(a0)
+	lbu	t1, 0(a1)
+	bne	t0, t1, 3f
+	addi	a0, a0, 1
+	addi	a1, a1, 1
+	bne	a0, a2, 1b
+2:
+	li	a0, 0
+	ret
+3:
+	sub	a0, t0, t1
+	ret
+
+
+memcmp_zbb:
+.option push
+.option arch,+zbb
+/*
+ * 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
+ */
+	beq	a0, a1, 4f
+
+	li	t0, SZREG
+	bltu	a2, t0, 5f
+
+1:
+	REG_L	t1, 0(a0)
+	REG_L	t2, 0(a1)
+	bne	t1, t2, 2f
+
+	addi	a0, a0, SZREG
+	addi	a1, a1, SZREG
+	addi	a2, a2, -SZREG
+	bgeu	a2, t0, 1b
+
+5:
+	beqz	a2, 4f
+6:
+	lbu	t1, 0(a0)
+	lbu	t2, 0(a1)
+	bne	t1, t2, 3f
+	addi	a0, a0, 1
+	addi	a1, a1, 1
+	addi	a2, a2, -1
+	bnez	a2, 6b
+
+4:	li	a0, 0
+	ret
+2:
+#ifndef CONFIG_CPU_BIG_ENDIAN
+	rev8	t1, t1
+	rev8	t2, t2
+#endif
+	sltu	a0, t2, t1
+	sltu	t0, t1, t2
+	sub	a0, a0, t0
+	ret
+
+3:
+	sub	a0, t1, t2
+	ret
+
+.option pop
+
+SYM_FUNC_END(memcmp)
+SYM_FUNC_ALIAS(__pi_memcmp, memcmp)
+EXPORT_SYMBOL(memcmp)
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 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>
---
 lib/tests/string_kunit.c | 102 +++++++++++++++++++++++++++++++++++++++
 1 file changed, 102 insertions(+)

diff --git a/lib/tests/string_kunit.c b/lib/tests/string_kunit.c
index 0819ace5b..7b4aa6eee 100644
--- a/lib/tests/string_kunit.c
+++ b/lib/tests/string_kunit.c
@@ -880,6 +880,106 @@ 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 int max_offset = 16;
+	const int max_len = 32;
+	const int buf_size = max_offset + max_len + 32;
+	u8 *buf1, *buf2;
+	int i, j, len, k;
+
+	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:%d j:%d len:%d", i, j, len);
+				for (k = 0; k < len; k++) {
+					memset(buf1, 'A', buf_size);
+					memset(buf2, 'A', buf_size);
+					buf2[j + k] = 'B';
+					int res = memcmp(buf1 + i, buf2 + j, len);
+
+					KUNIT_EXPECT_NE_MSG(test, res, 0,
+						"Should detect difference at k:%d (i:%d j:%d len:%d)",
+						k, i, j, len);
+
+					if (buf1[i + k] < buf2[j + k])
+						KUNIT_EXPECT_LT(test, res, 0);
+					else
+						KUNIT_EXPECT_GT(test, res, 0);
+				}
+			}
+		}
+	}
+}
+#if IS_ENABLED(CONFIG_STRING_KUNIT_BENCH)
+static void string_bench_memcmp(struct kunit *test)
+{
+	char *buf1, *buf2;
+	size_t lengths[] = { 1, 7, 8, 16, 32, 64, 128, 512, 1024, 4096};
+	int offsets[] = {0, 1, 3, 7};
+	const size_t max_len = 4096 + 64;
+	int res = 0;
+
+	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 (int i = 0; i < 100000; i++)
+		res += memcmp(buf1, buf2, 4096);
+
+	for (int o = 0; o < ARRAY_SIZE(offsets); o++) {
+		int off = offsets[o];
+
+		for (int i = 0; i < ARRAY_SIZE(lengths); i++) {
+			size_t len = lengths[i];
+			char *p1 = buf1;
+			char *p2 = buf2 + off;
+
+			u32 iterations = (len < 512) ? 100000 : 10000;
+
+			for (u32 j = 0; j < iterations; j++) {
+				asm volatile("" : : "g"(p1), "g"(p2), "g"(len) : "memory");
+				int res = memcmp(p1, p2, len);
+
+				asm volatile("" : : "g"(res) : "memory");
+			}
+
+			u64 elapsed = STRING_BENCH(iterations, memcmp, p1, p2, len);
+			u64 ns_per_call = div_u64(elapsed, iterations);
+			u64 mbps = len ? div_u64((u64)len * iterations * 1000,
+									 elapsed) : 0;
+
+			if (off == 0) {
+				kunit_info(test, "bench_memcmp_aligned: len=%-4zu: %llu MB/s (%llu ns/call)\n",
+					   len, mbps, ns_per_call);
+			} else {
+				kunit_info(test, "bench_memcmp_unaligned(off=%d): len=%-4zu: %llu MB/s (%llu ns/call)\n",
+					   off, len, mbps, ns_per_call);
+			}
+		}
+	}
+
+	vfree(buf1);
+	vfree(buf2);
+}
+#endif
 
 static struct kunit_case string_test_cases[] = {
 	KUNIT_CASE(string_test_memset16),
@@ -910,6 +1010,8 @@ 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 1/2] riscv: lib: add memcmp() implementation

[David Laight]

On Tue, 12 May 2026 16:10:06 +0200
Milan Tripkovic <milant2002@gmail.com> wrote:

> From: Milan Tripkovic <Milan.Tripkovic@rt-rk.com>
> 
> Add an assembly implementation of memcmp() for RISC-V. The implementation
> uses the ZBB extension for word-at-a-time comparison and an assembly
> fallback for non-ZBB systems.
> 
> Benchmark results (QEMU TCG, rv64):
> 
>   Len  | Def   | NoZBB | ZBB   | %NoZBB | %ZBB
>   -----|-------|-------|-------|--------|-------
>   1 B  | 22.4  | 24.6  | 23.2  | +9.8%  | +3.5%
>   7 B  | 96.9  | 108.5 | 107.3 | +12.0% | +10.7%
>   8 B  | 107.0 | 116.3 | 176.7 | +8.7%  | +65.1%
>   16 B | 148.4 | 172.8 | 315.6 | +16.4% | +112.6%
>   31 B | 182.2 | 217.1 | 377.6 | +19.2% | +107.2%
>   64 B | 220.6 | 239.4 | 874.2 | +8.5%  | +296.2%
>   127 B| 213.7 | 254.8 | 1042.9| +19.2% | +388.0%
>   512 B| 255.1 | 269.0 | 1778.6| +5.4%  | +597.2%
>   1024B| 252.3 | 280.9 | 1887.7| +11.3% | +648.1%
>   3173B| 241.3 | 288.7 | 2063.2| +19.6% | +755.0%
>   4096B| 240.9 | 280.5 | 2064.5| +16.4% | +756.9%
> 
> Signed-off-by: Milan Tripkovic <Milan.Tripkovic@rt-rk.com>
> ---
>  arch/riscv/include/asm/string.h |   2 +
>  arch/riscv/lib/Makefile         |   1 +
>  arch/riscv/lib/memcmp.S         | 103 ++++++++++++++++++++++++++++++++
>  arch/riscv/purgatory/Makefile   |   5 +-
>  4 files changed, 110 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..444b082d9
> --- /dev/null
> +++ b/arch/riscv/lib/memcmp.S
> @@ -0,0 +1,103 @@
> +/* SPDX-License-Identifier: GPL-2.0-only */
> +
> +#include <linux/linkage.h>
> +#include <asm/asm.h>
> +#include <asm/alternative-macros.h>
> +#include <asm/hwcap.h>
> +
> +/* int memcmp(const void *cs, const void *ct, size_t n) */
> +SYM_FUNC_START(memcmp)
> +
> +	__ALTERNATIVE_CFG("nop", "j memcmp_zbb", 0, RISCV_ISA_EXT_ZBB,
> +		IS_ENABLED(CONFIG_RISCV_ISA_ZBB) && IS_ENABLED(CONFIG_TOOLCHAIN_HAS_ZBB))
> +/*
> + * 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), transformed to end pointer (a0 + n)
> + *
> + * Returns
> + *	a0 - 0 if equal, positive if cs > ct, negative if cs < ct
> + *
> + * Clobbers
> + *	t0, t1
> + */
> +	beqz	a2, 2f
> +	add	a2, a0, a2
> +1:
> +	lbu	t0, 0(a0)
> +	lbu	t1, 0(a1)
> +	bne	t0, t1, 3f
> +	addi	a0, a0, 1
> +	addi	a1, a1, 1
> +	bne	a0, a2, 1b
> +2:
> +	li	a0, 0
> +	ret
> +3:
> +	sub	a0, t0, t1
> +	ret
> +
> +
> +memcmp_zbb:
> +.option push
> +.option arch,+zbb
> +/*
> + * 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
> + */
> +	beq	a0, a1, 4f

There is no point optimising for equal pointers.

> +
> +	li	t0, SZREG
> +	bltu	a2, t0, 5f
> +
> +1:
> +	REG_L	t1, 0(a0)
> +	REG_L	t2, 0(a1)

Aren't there some systems where misaligned reads are very expensive?
You might want to fall back to byte compares for misaligned buffers.

> +	bne	t1, t2, 2f
> +
> +	addi	a0, a0, SZREG
> +	addi	a1, a1, SZREG
> +	addi	a2, a2, -SZREG
> +	bgeu	a2, t0, 1b

You've a loop with two comparisons it in.
Move the length one to the top and the check before the loop shouldn't
be needed.

If you calculate the end address of one of the buffers you only
need two increments in the loop, not three.
You might need to access -SZREG(a0) to get the data.

> +
> +5:
> +	beqz	a2, 4f

If a0 and a1 are aligned you can read the next full word,
shift right (LE, left BE) and then compare.

> +6:
> +	lbu	t1, 0(a0)
> +	lbu	t2, 0(a1)
> +	bne	t1, t2, 3f
> +	addi	a0, a0, 1
> +	addi	a1, a1, 1
> +	addi	a2, a2, -1
> +	bnez	a2, 6b
> +
> +4:	li	a0, 0
> +	ret
> +2:
> +#ifndef CONFIG_CPU_BIG_ENDIAN
> +	rev8	t1, t1
> +	rev8	t2, t2
> +#endif

That looks like the only bit that needs zbb?
Is BIG_ENDIAN common enough to actually worry about?
You could just fall back to byte accesses (rereading memory) on BE.

-- David

> +	sltu	a0, t2, t1
> +	sltu	t0, t1, t2
> +	sub	a0, a0, t0
> +	ret
> +
> +3:
> +	sub	a0, t1, t2
> +	ret
> +
> +.option pop
> +
> +SYM_FUNC_END(memcmp)
> +SYM_FUNC_ALIAS(__pi_memcmp, memcmp)
> +EXPORT_SYMBOL(memcmp)
> 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