Re: [RFC PATCH 16/25] kvx: Add some library functions

From: Clément Léger
Date: Thu Jan 05 2023 - 08:04:09 EST


Le Tue, 3 Jan 2023 17:43:50 +0100,
Yann Sionneau <ysionneau@xxxxxxxxx> a écrit :

> Add some library functions for kvx, including: delay, memset,
> memcpy, strlen, clear_page, copy_page, raw_copy_from/to_user,
> asm_clear_user.
>
> CC: linux-kernel@xxxxxxxxxxxxxxx
> Co-developed-by: Clement Leger <clement.leger@xxxxxxxxxxx>
> Signed-off-by: Clement Leger <clement.leger@xxxxxxxxxxx>
> Co-developed-by: Jules Maselbas <jmaselbas@xxxxxxxxx>
> Signed-off-by: Jules Maselbas <jmaselbas@xxxxxxxxx>
> Co-developed-by: Julian Vetter <jvetter@xxxxxxxxx>
> Signed-off-by: Julian Vetter <jvetter@xxxxxxxxx>
> Co-developed-by: Marius Gligor <mgligor@xxxxxxxxx>
> Signed-off-by: Marius Gligor <mgligor@xxxxxxxxx>
> Co-developed-by: Yann Sionneau <ysionneau@xxxxxxxxx>
> Signed-off-by: Yann Sionneau <ysionneau@xxxxxxxxx>
> ---
> arch/kvx/include/asm/string.h | 20 ++
> arch/kvx/kernel/kvx_ksyms.c | 5 +
> arch/kvx/lib/clear_page.S | 40 ++++
> arch/kvx/lib/copy_page.S | 90 +++++++++
> arch/kvx/lib/delay.c | 39 ++++
> arch/kvx/lib/memcpy.c | 70 +++++++
> arch/kvx/lib/memset.S | 351 ++++++++++++++++++++++++++++++++++
> arch/kvx/lib/strlen.S | 122 ++++++++++++
> arch/kvx/lib/usercopy.S | 90 +++++++++
> 9 files changed, 827 insertions(+)
> create mode 100644 arch/kvx/include/asm/string.h
> create mode 100644 arch/kvx/lib/clear_page.S
> create mode 100644 arch/kvx/lib/copy_page.S
> create mode 100644 arch/kvx/lib/delay.c
> create mode 100644 arch/kvx/lib/memcpy.c
> create mode 100644 arch/kvx/lib/memset.S
> create mode 100644 arch/kvx/lib/strlen.S
> create mode 100644 arch/kvx/lib/usercopy.S
>
> diff --git a/arch/kvx/include/asm/string.h b/arch/kvx/include/asm/string.h
> new file mode 100644
> index 000000000000..677c1393a5cd
> --- /dev/null
> +++ b/arch/kvx/include/asm/string.h
> @@ -0,0 +1,20 @@
> +/* SPDX-License-Identifier: GPL-2.0-only */
> +/*
> + * Copyright (C) 2017-2023 Kalray Inc.
> + * Author(s): Clement Leger
> + * Jules Maselbas
> + */
> +
> +#ifndef _ASM_KVX_STRING_H
> +#define _ASM_KVX_STRING_H
> +
> +#define __HAVE_ARCH_MEMSET
> +extern void *memset(void *s, int c, size_t n);
> +
> +#define __HAVE_ARCH_MEMCPY
> +extern void *memcpy(void *dest, const void *src, size_t n);
> +
> +#define __HAVE_ARCH_STRLEN
> +extern size_t strlen(const char *s);
> +
> +#endif /* _ASM_KVX_STRING_H */
> diff --git a/arch/kvx/kernel/kvx_ksyms.c b/arch/kvx/kernel/kvx_ksyms.c
> index 18990aaf259f..678f81716dea 100644
> --- a/arch/kvx/kernel/kvx_ksyms.c
> +++ b/arch/kvx/kernel/kvx_ksyms.c
> @@ -22,3 +22,8 @@ DECLARE_EXPORT(__umoddi3);
> DECLARE_EXPORT(__divdi3);
> DECLARE_EXPORT(__udivdi3);
> DECLARE_EXPORT(__multi3);
> +
> +DECLARE_EXPORT(clear_page);
> +DECLARE_EXPORT(copy_page);
> +DECLARE_EXPORT(memset);
> +DECLARE_EXPORT(asm_clear_user);
> diff --git a/arch/kvx/lib/clear_page.S b/arch/kvx/lib/clear_page.S
> new file mode 100644
> index 000000000000..364fe0663ca2
> --- /dev/null
> +++ b/arch/kvx/lib/clear_page.S
> @@ -0,0 +1,40 @@
> +/* SPDX-License-Identifier: GPL-2.0-only */
> +/*
> + * Copyright (C) 2017-2023 Kalray Inc.
> + * Author(s): Marius Gligor
> + * Clement Leger
> + */
> +
> +#include <linux/linkage.h>
> +#include <linux/export.h>
> +#include <linux/const.h>
> +
> +#include <asm/cache.h>
> +#include <asm/page.h>
> +
> +#define CLEAR_PAGE_LOOP_COUNT (PAGE_SIZE / 32)
> +
> +/*
> + * Clear page @dest.
> + *
> + * Parameters:
> + * r0 - dest page
> + */
> +ENTRY(clear_page)
> + make $r1 = CLEAR_PAGE_LOOP_COUNT
> + ;;
> + make $r4 = 0
> + make $r5 = 0
> + make $r6 = 0
> + make $r7 = 0
> + ;;
> +
> + loopdo $r1, clear_page_done
> + ;;
> + so 0[$r0] = $r4r5r6r7
> + addd $r0 = $r0, 32
> + ;;
> + clear_page_done:
> + ret
> + ;;
> +ENDPROC(clear_page)
> diff --git a/arch/kvx/lib/copy_page.S b/arch/kvx/lib/copy_page.S
> new file mode 100644
> index 000000000000..4bb82d1c964c
> --- /dev/null
> +++ b/arch/kvx/lib/copy_page.S
> @@ -0,0 +1,90 @@
> +/* SPDX-License-Identifier: GPL-2.0-only */
> +/*
> + * Copyright (C) 2017-2023 Kalray Inc.
> + * Author(s): Clement Leger
> + */
> +
> +#include <linux/linkage.h>
> +#include <linux/const.h>
> +
> +#include <asm/page.h>
> +
> +/* We have 8 load/store octuple (32 bytes) per hardware loop */
> +#define COPY_SIZE_PER_LOOP (32 * 8)
> +#define COPY_PAGE_LOOP_COUNT (PAGE_SIZE / COPY_SIZE_PER_LOOP)
> +
> +/*
> + * Copy a page from src to dest (both are page aligned)
> + * In order to recover from smem latency, unroll the loop to trigger multiple
> + * onfly loads and avoid waiting too much for them to return.
> + * We use 8 * 32 load even though we could use more (up to 10 loads) to simplify
> + * the handling using a single hardware loop
> + *
> + * Parameters:
> + * r0 - dest
> + * r1 - src
> + */
> +ENTRY(copy_page)
> + make $r2 = COPY_PAGE_LOOP_COUNT
> + make $r3 = 0
> + ;;
> + loopdo $r2, copy_page_done
> + ;;
> + /*
> + * Load 8 * 32 bytes using uncached access to avoid hitting
> + * the cache
> + */
> + lo.xs $r32r33r34r35 = $r3[$r1]
> + /* Copy current copy index for store */
> + copyd $r2 = $r3
> + addd $r3 = $r3, 1
> + ;;
> + lo.xs $r36r37r38r39 = $r3[$r1]
> + addd $r3 = $r3, 1
> + ;;
> + lo.xs $r40r41r42r43 = $r3[$r1]
> + addd $r3 = $r3, 1
> + ;;
> + lo.xs $r44r45r46r47 = $r3[$r1]
> + addd $r3 = $r3, 1
> + ;;
> + lo.xs $r48r49r50r51 = $r3[$r1]
> + addd $r3 = $r3, 1
> + ;;
> + lo.xs $r52r53r54r55 = $r3[$r1]
> + addd $r3 = $r3, 1
> + ;;
> + lo.xs $r56r57r58r59 = $r3[$r1]
> + addd $r3 = $r3, 1
> + ;;
> + lo.xs $r60r61r62r63 = $r3[$r1]
> + addd $r3 = $r3, 1
> + ;;
> + /* And then store all of them */
> + so.xs $r2[$r0] = $r32r33r34r35
> + addd $r2 = $r2, 1
> + ;;
> + so.xs $r2[$r0] = $r36r37r38r39
> + addd $r2 = $r2, 1
> + ;;
> + so.xs $r2[$r0] = $r40r41r42r43
> + addd $r2 = $r2, 1
> + ;;
> + so.xs $r2[$r0] = $r44r45r46r47
> + addd $r2 = $r2, 1
> + ;;
> + so.xs $r2[$r0] = $r48r49r50r51
> + addd $r2 = $r2, 1
> + ;;
> + so.xs $r2[$r0] = $r52r53r54r55
> + addd $r2 = $r2, 1
> + ;;
> + so.xs $r2[$r0] = $r56r57r58r59
> + addd $r2 = $r2, 1
> + ;;
> + so.xs $r2[$r0] = $r60r61r62r63
> + ;;
> + copy_page_done:
> + ret
> + ;;
> +ENDPROC(copy_page)
> diff --git a/arch/kvx/lib/delay.c b/arch/kvx/lib/delay.c
> new file mode 100644
> index 000000000000..11295eedc3f5
> --- /dev/null
> +++ b/arch/kvx/lib/delay.c
> @@ -0,0 +1,39 @@
> +// SPDX-License-Identifier: GPL-2.0-only
> +/*
> + * Copyright (C) 2017-2023 Kalray Inc.
> + * Author(s): Clement Leger
> + */
> +
> +#include <linux/export.h>
> +#include <linux/delay.h>
> +
> +#include <asm/param.h>
> +#include <asm/timex.h>
> +
> +void __delay(unsigned long loops)
> +{
> + cycles_t target_cycle = get_cycles() + loops;
> +
> + while (get_cycles() < target_cycle);
> +}
> +EXPORT_SYMBOL(__delay);
> +
> +inline void __const_udelay(unsigned long xloops)
> +{
> + u64 loops = (u64)xloops * (u64)loops_per_jiffy * HZ;
> +
> + __delay(loops >> 32);
> +}
> +EXPORT_SYMBOL(__const_udelay);
> +
> +void __udelay(unsigned long usecs)
> +{
> + __const_udelay(usecs * 0x10C7UL); /* 2**32 / 1000000 (rounded up) */
> +}
> +EXPORT_SYMBOL(__udelay);

This could probably be factorized with arm64, csky, nios2, etc.

> +
> +void __ndelay(unsigned long nsecs)
> +{
> + __const_udelay(nsecs * 0x5UL); /* 2**32 / 1000000000 (rounded up) */
> +}
> +EXPORT_SYMBOL(__ndelay);

Same goes for __ndelay I think

> diff --git a/arch/kvx/lib/memcpy.c b/arch/kvx/lib/memcpy.c
> new file mode 100644
> index 000000000000..b81f746a80ee
> --- /dev/null
> +++ b/arch/kvx/lib/memcpy.c
> @@ -0,0 +1,70 @@
> +// SPDX-License-Identifier: GPL-2.0-only
> +/*
> + * Copyright (C) 2017-2023 Kalray Inc.
> + * Author(s): Clement Leger
> + * Yann Sionneau
> + */
> +
> +#include <linux/export.h>
> +#include <linux/types.h>
> +
> +void *memcpy(void *dest, const void *src, size_t n)
> +{
> + __uint128_t *tmp128_d = dest;
> + const __uint128_t *tmp128_s = src;
> + uint64_t *tmp64_d;
> + const uint64_t *tmp64_s;
> + uint32_t *tmp32_d;
> + const uint32_t *tmp32_s;
> + uint16_t *tmp16_d;
> + const uint16_t *tmp16_s;
> + uint8_t *tmp8_d;
> + const uint8_t *tmp8_s;
> +
> + while (n >= 16) {
> + *tmp128_d = *tmp128_s;
> + tmp128_d++;
> + tmp128_s++;
> + n -= 16;
> + }
> +
> + tmp64_d = (uint64_t *) tmp128_d;
> + tmp64_s = (uint64_t *) tmp128_s;
> + while (n >= 8) {
> + *tmp64_d = *tmp64_s;
> + tmp64_d++;
> + tmp64_s++;
> + n -= 8;
> + }
> +
> + tmp32_d = (uint32_t *) tmp64_d;
> + tmp32_s = (uint32_t *) tmp64_s;
> + while (n >= 4) {
> + *tmp32_d = *tmp32_s;
> + tmp32_d++;
> + tmp32_s++;
> + n -= 4;
> + }
> +
> + tmp16_d = (uint16_t *) tmp32_d;
> + tmp16_s = (uint16_t *) tmp32_s;
> + while (n >= 2) {
> + *tmp16_d = *tmp16_s;
> + tmp16_d++;
> + tmp16_s++;
> + n -= 2;
> + }
> +
> + tmp8_d = (uint8_t *) tmp16_d;
> + tmp8_s = (uint8_t *) tmp16_s;
> + while (n >= 1) {
> + *tmp8_d = *tmp8_s;
> + tmp8_d++;
> + tmp8_s++;
> + n--;
> + }
> +
> + return dest;
> +}
> +EXPORT_SYMBOL(memcpy);
> +
> diff --git a/arch/kvx/lib/memset.S b/arch/kvx/lib/memset.S
> new file mode 100644
> index 000000000000..9eebc28da2be
> --- /dev/null
> +++ b/arch/kvx/lib/memset.S
> @@ -0,0 +1,351 @@
> +/* SPDX-License-Identifier: GPL-2.0-only */
> +/*
> + * Copyright (C) 2017-2023 Kalray Inc.
> + * Author(s): Clement Leger
> + * Marius Gligor
> + */
> +
> +#include <linux/linkage.h>
> +
> +#include <asm/cache.h>
> +
> +#define REPLICATE_BYTE_MASK 0x0101010101010101
> +#define MIN_SIZE_FOR_ALIGN 128
> +
> +/*
> + * Optimized memset for kvx architecture
> + *
> + * In order to optimize memset on kvx, we can use various things:
> + * - conditionnal store which avoid branch penalty
> + * - store half/word/double/quad/octuple to store up to 16 bytes at a time
> + * - dzerol to zero a cacheline when the pattern is '0' (often the case)
> + * - hardware loop for steady cases.
> + *
> + * First, we assume that memset is mainly used for zeroing areas. In order
> + * to optimize this case, we consider it to be the fast path of the algorithm.
> + * In both cases (0 and non 0 pattern), we start by checking if the size is
> + * below a minimum size. If so, we skip the alignment part. Indeed, the kvx
> + * supports misalignment and the penalty for letting it do unaligned accesses is
> + * lower than trying to realigning us. So for small sizes, we don't even bother
> + * to realign. Minor difference is that in the memset with 0, we skip after the
> + * dzerol loop since dzerol must be cache-line aligned (no misalignment of
> + * course).
> + * Regarding the non 0 pattern memset, we use sbmm to replicate the pattern on
> + * all bits on a register in one call.
> + * Once alignment has been reached, we can do the hardware loop for both cases(
> + * store octuple/dzerol) in order to optimize throughput. Care must be taken to
> + * align hardware loops on at least 8 bytes for performances.
> + * Once the main loop has been done, we finish the copy by checking length to do
> + * the necessary calls to store remaining bytes.
> + *
> + * Pseudo code (applies for non 0 pattern):
> + *
> + * int memset(void *dest, char pattern, long length)
> + * {
> + * long dest_align = -((long) dest);
> + * long copy;
> + * long orig_dest = dest;
> + *
> + * uint64_t pattern = sbmm8(pattern, 0x0101010101010101);
> + * uint128_t pattern128 = pattern << 64 | pattern;
> + * uint256_t pattern128 = pattern128 << 128 | pattern128;
> + *
> + * // Keep only low bits
> + * dest_align &= 0x1F;
> + * length -= dest_align;
> + *
> + * // Byte align
> + * copy = align & (1 << 0);
> + * if (copy)
> + * *((u8 *) dest) = pattern;
> + * dest += copy;
> + * // Half align
> + * copy = align & (1 << 1);
> + * if (copy)
> + * *((u16 *) dest) = pattern;
> + * dest += copy;
> + * // Word align
> + * copy = align & (1 << 2);
> + * if (copy)
> + * *((u32 *) dest) = pattern;
> + * dest += copy;
> + * // Double align
> + * copy = align & (1 << 3);
> + * if (copy)
> + * *((u64 *) dest) = pattern;
> + * dest += copy;
> + * // Quad align
> + * copy = align & (1 << 4);
> + * if (copy)
> + * *((u128 *) dest) = pattern128;
> + * dest += copy;
> + *
> + * // We are now aligned on 256 bits
> + * loop_octuple_count = size >> 5;
> + * for (i = 0; i < loop_octuple_count; i++) {
> + * *((u256 *) dest) = pattern256;
> + * dest += 32;
> + * }
> + *
> + * if (length == 0)
> + * return orig_dest;
> + *
> + * // Copy remaining part
> + * remain = length & (1 << 4);
> + * if (copy)
> + * *((u128 *) dest) = pattern128;
> + * dest += remain;
> + * remain = length & (1 << 3);
> + * if (copy)
> + * *((u64 *) dest) = pattern;
> + * dest += remain;
> + * remain = length & (1 << 2);
> + * if (copy)
> + * *((u32 *) dest) = pattern;
> + * dest += remain;
> + * remain = length & (1 << 1);
> + * if (copy)
> + * *((u16 *) dest) = pattern;
> + * dest += remain;
> + * remain = length & (1 << 0);
> + * if (copy)
> + * *((u8 *) dest) = pattern;
> + * dest += remain;
> + *
> + * return orig_dest;
> + * }
> + */
> +
> +.text
> +.align 16
> +ENTRY(memset):
> + make $r32 = 0
> + make $r33 = 0
> + /* Check if length < KVX_DCACHE_LINE_SIZE */
> + compd.ltu $r7 = $r2, KVX_DCACHE_LINE_SIZE
> + /* Jump to generic memset if pattern is != 0 */
> + cb.dnez $r1? memset_non_0_pattern
> + ;;
> + /* Preserve return value */
> + copyd $r3 = $r0
> + /* Invert address to compute size to copy to be aligned on 32 bytes */
> + negd $r5 = $r0
> + /* Remaining bytes for 16 bytes store (for alignment on 64 bytes) */
> + andd $r8 = $r2, (1 << 5)
> + copyq $r34r35 = $r32, $r33
> + /* Skip loopdo with dzerol if length < KVX_DCACHE_LINE_SIZE */
> + cb.dnez $r7? .Ldzerol_done
> + ;;
> + /* Compute the size that will be copied to align on 64 bytes boundary */
> + andw $r6 = $r5, 0x3F
> + /* Check if address is aligned on 64 bytes */
> + andw $r9 = $r0, 0x3F
> + /* Alignment */
> + nop
> + ;;
> + /* If address already aligned on 64 bytes, jump to dzerol loop */
> + cb.deqz $r9? .Laligned_64
> + /* Remove unaligned part from length */
> + sbfd $r2 = $r6, $r2
> + /* Check if we need to copy 1 byte */
> + andw $r4 = $r5, (1 << 0)
> + ;;
> + /* If we are not aligned, store byte */
> + sb.dnez $r4? [$r0] = $r32
> + /* Check if we need to copy 2 bytes */
> + andw $r4 = $r5, (1 << 1)
> + /* Add potentially copied part for next store offset */
> + addd $r0 = $r0, $r4
> + ;;
> + sh.dnez $r4? [$r0] = $r32
> + /* Check if we need to copy 4 bytes */
> + andw $r4 = $r5, (1 << 2)
> + addd $r0 = $r0, $r4
> + ;;
> + sw.dnez $r4? [$r0] = $r32
> + /* Check if we need to copy 8 bytes */
> + andw $r4 = $r5, (1 << 3)
> + addd $r0 = $r0, $r4
> + ;;
> + sd.dnez $r4? [$r0] = $r32
> + /* Check if we need to copy 16 bytes */
> + andw $r4 = $r5, (1 << 4)
> + addd $r0 = $r0, $r4
> + ;;
> + sq.dnez $r4? [$r0] = $r32r33
> + /* Check if we need to copy 32 bytes */
> + andw $r4 = $r5, (1 << 5)
> + addd $r0 = $r0, $r4
> + ;;
> + so.dnez $r4? [$r0] = $r32r33r34r35
> + addd $r0 = $r0, $r4
> + ;;
> +.Laligned_64:
> + /* Prepare amount of data for dzerol */
> + srld $r10 = $r2, 6
> + /* Size to be handled in loopdo */
> + andd $r4 = $r2, ~0x3F
> + make $r11 = 64
> + cb.deqz $r2? .Lmemset_done
> + ;;
> + /* Remaining bytes for 16 bytes store */
> + andw $r8 = $r2, (1 << 5)
> + /* Skip dzerol if there are not enough data for 64 bytes store */
> + cb.deqz $r10? .Ldzerol_done
> + /* Update length to copy */
> + sbfd $r2 = $r4, $r2
> + ;;
> + loopdo $r10, .Ldzerol_done
> + ;;
> + so 0[$r0], $r32r33r34r35
> + ;;
> + so 32[$r0], $r32r33r34r35
> + addd $r0 = $r0, $r11
> + ;;
> + .Ldzerol_done:
> + /*
> + * Now that we have handled every aligned bytes using 'dzerol', we can
> + * handled the remainder of length using store by decrementing size
> + * We also exploit the fact we are aligned to simply check remaining
> + * size */
> + so.dnez $r8? [$r0] = $r32r33r34r35
> + addd $r0 = $r0, $r8
> + /* Remaining bytes for 16 bytes store */
> + andw $r8 = $r2, (1 << 4)
> + cb.deqz $r2? .Lmemset_done
> + ;;
> + sq.dnez $r8? [$r0] = $r32r33
> + addd $r0 = $r0, $r8
> + /* Remaining bytes for 8 bytes store */
> + andw $r8 = $r2, (1 << 3)
> + ;;
> + sd.dnez $r8? [$r0] = $r32
> + addd $r0 = $r0, $r8
> + /* Remaining bytes for 4 bytes store */
> + andw $r8 = $r2, (1 << 2)
> + ;;
> + sw.dnez $r8? [$r0] = $r32
> + addd $r0 = $r0, $r8
> + /* Remaining bytes for 2 bytes store */
> + andw $r8 = $r2, (1 << 1)
> + ;;
> + sh.dnez $r8? [$r0] = $r32
> + addd $r0 = $r0, $r8
> + ;;
> + sb.odd $r2? [$r0] = $r32
> + /* Restore original value */
> + copyd $r0 = $r3
> + ret
> + ;;
> +
> +.align 16
> +memset_non_0_pattern:
> + /* Preserve return value */
> + copyd $r3 = $r0
> + /* Replicate the first pattern byte on all bytes */
> + sbmm8 $r32 = $r1, REPLICATE_BYTE_MASK
> + /* Check if length < MIN_SIZE_FOR_ALIGN */
> + compd.geu $r7 = $r2, MIN_SIZE_FOR_ALIGN
> + /* Invert address to compute size to copy to be aligned on 32 bytes */
> + negd $r5 = $r0
> + ;;
> + /* Check if we are aligned on 32 bytes */
> + andw $r9 = $r0, 0x1F
> + /* Compute the size that will be copied to align on 32 bytes boundary */
> + andw $r6 = $r5, 0x1F
> + /*
> + * If size < MIN_SIZE_FOR_ALIGN bits, directly go to so, it will be done
> + * unaligned but that is still better that what we can do with sb
> + */
> + cb.deqz $r7? .Laligned_32
> + ;;
> + /* Remove unaligned part from length */
> + sbfd $r2 = $r6, $r2
> + /* If we are already aligned on 32 bytes, jump to main "so" loop */
> + cb.deqz $r9? .Laligned_32
> + /* Check if we need to copy 1 byte */
> + andw $r4 = $r5, (1 << 0)
> + ;;
> + /* If we are not aligned, store byte */
> + sb.dnez $r4? [$r0] = $r32
> + /* Check if we need to copy 2 bytes */
> + andw $r4 = $r5, (1 << 1)
> + /* Add potentially copied part for next store offset */
> + addd $r0 = $r0, $r4
> + ;;
> + sh.dnez $r4? [$r0] = $r32
> + /* Check if we need to copy 4 bytes */
> + andw $r4 = $r5, (1 << 2)
> + addd $r0 = $r0, $r4
> + ;;
> + sw.dnez $r4? [$r0] = $r32
> + /* Check if we need to copy 8 bytes */
> + andw $r4 = $r5, (1 << 3)
> + addd $r0 = $r0, $r4
> + /* Copy second part of pattern for sq */
> + copyd $r33 = $r32
> + ;;
> + sd.dnez $r4? [$r0] = $r32
> + /* Check if we need to copy 16 bytes */
> + andw $r4 = $r5, (1 << 4)
> + addd $r0 = $r0, $r4
> + ;;
> + sq.dnez $r4? [$r0] = $r32r33
> + addd $r0 = $r0, $r4
> + ;;
> +.Laligned_32:
> + /* Copy second part of pattern for sq */
> + copyd $r33 = $r32
> + /* Prepare amount of data for 32 bytes store */
> + srld $r10 = $r2, 5
> + nop
> + nop
> + ;;
> + copyq $r34r35 = $r32, $r33
> + /* Remaining bytes for 16 bytes store */
> + andw $r8 = $r2, (1 << 4)
> + make $r11 = 32
> + /* Check if there are enough data for 32 bytes store */
> + cb.deqz $r10? .Laligned_32_done
> + ;;
> + loopdo $r10, .Laligned_32_done
> + ;;
> + so 0[$r0] = $r32r33r34r35
> + addd $r0 = $r0, $r11
> + ;;
> + .Laligned_32_done:
> + /*
> + * Now that we have handled every aligned bytes using 'so', we can
> + * handled the remainder of length using store by decrementing size
> + * We also exploit the fact we are aligned to simply check remaining
> + * size */
> + sq.dnez $r8? [$r0] = $r32r33
> + addd $r0 = $r0, $r8
> + /* Remaining bytes for 8 bytes store */
> + andw $r8 = $r2, (1 << 3)
> + cb.deqz $r2? .Lmemset_done
> + ;;
> + sd.dnez $r8? [$r0] = $r32
> + addd $r0 = $r0, $r8
> + /* Remaining bytes for 4 bytes store */
> + andw $r8 = $r2, (1 << 2)
> + ;;
> + sw.dnez $r8? [$r0] = $r32
> + addd $r0 = $r0, $r8
> + /* Remaining bytes for 2 bytes store */
> + andw $r8 = $r2, (1 << 1)
> + ;;
> + sh.dnez $r8? [$r0] = $r32
> + addd $r0 = $r0, $r8
> + ;;
> + sb.odd $r2? [$r0] = $r32
> + /* Restore original value */
> + copyd $r0 = $r3
> + ret
> + ;;
> +.Lmemset_done:
> + /* Restore original value */
> + copyd $r0 = $r3
> + ret
> + ;;
> +ENDPROC(memset)
> diff --git a/arch/kvx/lib/strlen.S b/arch/kvx/lib/strlen.S
> new file mode 100644
> index 000000000000..8298402a7898
> --- /dev/null
> +++ b/arch/kvx/lib/strlen.S
> @@ -0,0 +1,122 @@
> +/* SPDX-License-Identifier: GPL-2.0-only */
> +/*
> + * Copyright (C) 2017-2023 Kalray Inc.
> + * Author(s): Jules Maselbas
> + */
> +#include <linux/linkage.h>
> +#include <asm/export.h>
> +
> +/*
> + * kvx optimized strlen
> + *
> + * This implementation of strlen only does aligned memory accesses.
> + * Since we don't know the total length the idea is to do double word
> + * load and stop on the first null byte found. As it's always safe to
> + * read more up to a lower 8-bytes boundary.
> + *
> + * This implementation of strlen uses a trick to detect if a double
> + * word contains a null byte [1]:
> + *
> + * > #define haszero(v) (((v) - 0x01010101UL) & ~(v) & 0x80808080UL)
> + * > The sub-expression (v - 0x01010101UL), evaluates to a high bit set
> + * > in any byte whenever the corresponding byte in v is zero or greater
> + * > than 0x80. The sub-expression ~v & 0x80808080UL evaluates to high
> + * > bits set in bytes where the byte of v doesn't have its high bit set
> + * > (so the byte was less than 0x80). Finally, by ANDing these two sub-
> + * > expressions the result is the high bits set where the bytes in v
> + * > were zero, since the high bits set due to a value greater than 0x80
> + * > in the first sub-expression are masked off by the second.
> + *
> + * [1] http://graphics.stanford.edu/~seander/bithacks.html#ZeroInWord
> + *
> + * A second trick is used to get the exact number of characters before
> + * the first null byte in a double word:
> + *
> + * clz(sbmmt(zero, 0x0102040810204080))
> + *
> + * This trick uses the haszero result which maps null byte to 0x80 and
> + * others value to 0x00. The idea is to count the number of consecutive
> + * null byte in the double word (counting from less significant byte
> + * to most significant byte). To do so, using the bit matrix transpose
> + * will "pack" all high bit (0x80) to the most significant byte (MSB).
> + * It is not possible to count the trailing zeros in this MSB, however
> + * if a byte swap is done before the bit matrix transpose we still have
> + * all the information in the MSB but now we can count the leading zeros.
> + * The instruction sbmmt with the matrix 0x0102040810204080 does exactly
> + * what we need a byte swap followed by a bit transpose.
> + *
> + * A last trick is used to handle the first double word misalignment.
> + * This is done by masking off the N lower bytes (excess read) with N
> + * between 0 and 7. The mask is applied on haszero results and will
> + * force the N lower bytes to be considered not null.
> + *
> + * This is a C implementation of the algorithm described above:
> + *
> + * size_t strlen(char *s) {
> + * uint64_t *p = (uint64_t *)((uintptr_t)s) & ~0x7;
> + * uint64_t rem = ((uintptr_t)s) % 8;
> + * uint64_t low = -0x0101010101010101;
> + * uint64_t high = 0x8080808080808080;
> + * uint64_t dword, zero;
> + * uint64_t msk, len;
> + *
> + * dword = *p++;
> + * zero = (dword + low) & ~dword & high;
> + * msk = 0xffffffffffffffff << (rem * 8);
> + * zero &= msk;
> + *
> + * while (!zero) {
> + * dword = *p++;
> + * zero = (dword + low) & ~dword & high;
> + * }
> + *
> + * zero = __builtin_kvx_sbmmt8(zero, 0x0102040810204080);
> + * len = ((void *)p - (void *)s) - 8;
> + * len += __builtin_kvx_clzd(zero);
> + *
> + * return len;
> + * }
> + */
> +
> +.text
> +.align 16
> +ENTRY(strlen)
> + andd $r1 = $r0, ~0x7
> + andd $r2 = $r0, 0x7
> + make $r10 = -0x0101010101010101
> + make $r11 = 0x8080808080808080
> + ;;
> + ld $r4 = 0[$r1]
> + sllw $r2 = $r2, 3
> + make $r3 = 0xffffffffffffffff
> + ;;
> + slld $r2 = $r3, $r2
> + addd $r5 = $r4, $r10
> + andnd $r6 = $r4, $r11
> + ;;
> + andd $r6 = $r6, $r2
> + make $r3 = 0
> + ;;
> +.loop:
> + andd $r4 = $r5, $r6
> + addd $r1 = $r1, 0x8
> + ;;
> + cb.dnez $r4? .end
> + ld.deqz $r4? $r4 = [$r1]
> + ;;
> + addd $r5 = $r4, $r10
> + andnd $r6 = $r4, $r11
> + goto .loop
> + ;;
> +.end:
> + addd $r1 = $r1, -0x8
> + sbmmt8 $r4 = $r4, 0x0102040810204080
> + ;;
> + clzd $r4 = $r4
> + sbfd $r1 = $r0, $r1
> + ;;
> + addd $r0 = $r4, $r1
> + ret
> + ;;
> +ENDPROC(strlen)
> +EXPORT_SYMBOL(strlen)
> diff --git a/arch/kvx/lib/usercopy.S b/arch/kvx/lib/usercopy.S
> new file mode 100644
> index 000000000000..bc7e1a45e1c7
> --- /dev/null
> +++ b/arch/kvx/lib/usercopy.S
> @@ -0,0 +1,90 @@
> +/* SPDX-License-Identifier: GPL-2.0-only */
> +/*
> + * Copyright (C) 2017-2023 Kalray Inc.
> + * Author(s): Clement Leger
> + */
> +#include <linux/linkage.h>
> +
> +/**
> + * Copy from/to a user buffer
> + * r0 = to buffer
> + * r1 = from buffer
> + * r2 = size to copy
> + * This function can trapped when hitting a non-mapped page.
> + * It will trigger a trap NOMAPPING and the trap handler will interpret
> + * it and check if instruction pointer is inside __ex_table.
> + * The next step are described later !
> + */
> +.text
> +ENTRY(raw_copy_from_user)
> +ENTRY(raw_copy_to_user)
> + /**
> + * naive implementation byte per byte
> + */
> + make $r33 = 0x0;
> + /* If size == 0, exit directly */
> + cb.deqz $r2? copy_exit
> + ;;
> + loopdo $r2, copy_exit
> + ;;
> +0: lbz $r34 = $r33[$r1]
> + ;;
> +1: sb $r33[$r0] = $r34
> + addd $r33 = $r33, 1 /* Ptr increment */
> + addd $r2 = $r2, -1 /* Remaining bytes to copy */
> + ;;
> + copy_exit:
> + copyd $r0 = $r2
> + ret
> + ;;
> +ENDPROC(raw_copy_to_user)
> +ENDPROC(raw_copy_from_user)
> +
> +/**
> + * Exception table
> + * each entry correspond to the following:
> + * .dword trapping_addr, restore_addr
> + *
> + * On trap, the handler will try to locate if $spc is matching a
> + * trapping address in the exception table. If so, the restore addr
> + * will be put in the return address of the trap handler, allowing
> + * to properly finish the copy and return only the bytes copied/cleared
> + */
> +.pushsection __ex_table,"a"
> +.balign 8
> +.dword 0b, copy_exit
> +.dword 1b, copy_exit
> +.popsection
> +
> +/**
> + * Clear a user buffer
> + * r0 = buffer to clear
> + * r1 = size to clear
> + */
> +.text
> +ENTRY(asm_clear_user)
> + /**
> + * naive implementation byte per byte
> + */
> + make $r33 = 0x0;
> + make $r34 = 0x0;
> + /* If size == 0, exit directly */
> + cb.deqz $r1? clear_exit
> + ;;
> + loopdo $r1, clear_exit
> + ;;
> +40: sb $r33[$r0] = $r34
> + addd $r33 = $r33, 1 /* Ptr increment */
> + addd $r1 = $r1, -1 /* Remaining bytes to copy */
> + ;;
> + clear_exit:
> + copyd $r0 = $r1
> + ret
> + ;;
> +ENDPROC(asm_clear_user)
> +
> +.pushsection __ex_table,"a"
> +.balign 8
> +.dword 40b, clear_exit
> +.popsection
> +



--
Clément Léger,
Embedded Linux and Kernel engineer at Bootlin
https://bootlin.com