[arm] Add optimised string functions for 64-bit ARM

Signed-off-by: Michael Brown <mcb30@ipxe.org>

3 files changed, 355 insertions, 0 deletions

diff --git a/src/arch/arm/include/bits/string.h b/src/arch/arm32/include/bits/string.h
index 5b1c1505..5b1c1505 100644
--- a/src/arch/arm/include/bits/string.h
+++ b/src/arch/arm32/include/bits/string.h
diff --git a/src/arch/arm64/core/arm64_string.c b/src/arch/arm64/core/arm64_string.c
new file mode 100644
index 00000000..28a2b73b
--- /dev/null
+++ b/src/arch/arm64/core/arm64_string.c
@@ -0,0 +1,249 @@
+/*
+ * Copyright (C) 2016 Michael Brown <mbrown@fensystems.co.uk>.
+ *
+ * This program is free software; you can redistribute it and/or
+ * modify it under the terms of the GNU General Public License as
+ * published by the Free Software Foundation; either version 2 of the
+ * License, or any later version.
+ *
+ * This program is distributed in the hope that it will be useful, but
+ * WITHOUT ANY WARRANTY; without even the implied warranty of
+ * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the GNU
+ * General Public License for more details.
+ *
+ * You should have received a copy of the GNU General Public License
+ * along with this program; if not, write to the Free Software
+ * Foundation, Inc., 51 Franklin Street, Fifth Floor, Boston, MA
+ * 02110-1301, USA.
+ *
+ * You can also choose to distribute this program under the terms of
+ * the Unmodified Binary Distribution Licence (as given in the file
+ * COPYING.UBDL), provided that you have satisfied its requirements.
+ */
+
+/** @file
+ *
+ * Optimised string operations
+ *
+ */
+
+FILE_LICENCE ( GPL2_OR_LATER_OR_UBDL );
+
+#include <string.h>
+
+/**
+ * Copy memory area
+ *
+ * @v dest		Destination address
+ * @v src		Source address
+ * @v len		Length
+ * @ret dest		Destination address
+ */
+void arm64_memcpy ( void *dest, const void *src, size_t len ) {
+	void *discard_dest;
+	void *discard_end;
+	const void *discard_src;
+	size_t discard_offset;
+	unsigned long discard_data;
+	unsigned long discard_low;
+	unsigned long discard_high;
+
+	/* If length is too short for an "ldp"/"stp" instruction pair,
+	 * then just copy individual bytes.
+	 */
+	if ( len < 16 ) {
+		__asm__ __volatile__ ( "cbz %0, 2f\n\t"
+				       "\n1:\n\t"
+				       "sub %0, %0, #1\n\t"
+				       "ldrb %w1, [%3, %0]\n\t"
+				       "strb %w1, [%2, %0]\n\t"
+				       "cbnz %0, 1b\n\t"
+				       "\n2:\n\t"
+				       : "=&r" ( discard_offset ),
+					 "=&r" ( discard_data )
+				       : "r" ( dest ), "r" ( src ), "0" ( len )
+				       : "memory" );
+		return;
+	}
+
+	/* Use "ldp"/"stp" to copy 16 bytes at a time: one initial
+	 * potentially unaligned access, multiple destination-aligned
+	 * accesses, one final potentially unaligned access.
+	 */
+	__asm__ __volatile__ ( "ldp %3, %4, [%1], #16\n\t"
+			       "stp %3, %4, [%0], #16\n\t"
+			       "and %3, %0, #15\n\t"
+			       "sub %0, %0, %3\n\t"
+			       "sub %1, %1, %3\n\t"
+			       "bic %2, %5, #15\n\t"
+			       "b 2f\n\t"
+			       "\n1:\n\t"
+			       "ldp %3, %4, [%1], #16\n\t"
+			       "stp %3, %4, [%0], #16\n\t"
+			       "\n2:\n\t"
+			       "cmp %0, %2\n\t"
+			       "bne 1b\n\t"
+			       "ldp %3, %4, [%6, #-16]\n\t"
+			       "stp %3, %4, [%5, #-16]\n\t"
+			       : "=&r" ( discard_dest ),
+				 "=&r" ( discard_src ),
+				 "=&r" ( discard_end ),
+				 "=&r" ( discard_low ),
+				 "=&r" ( discard_high )
+			       : "r" ( dest + len ), "r" ( src + len ),
+				 "0" ( dest ), "1" ( src )
+			       : "memory", "cc" );
+}
+
+/**
+ * Zero memory region
+ *
+ * @v dest		Destination region
+ * @v len		Length
+ */
+void arm64_bzero ( void *dest, size_t len ) {
+	size_t discard_offset;
+	void *discard_dest;
+	void *discard_end;
+
+	/* If length is too short for an "stp" instruction, then just
+	 * zero individual bytes.
+	 */
+	if ( len < 16 ) {
+		__asm__ __volatile__ ( "cbz %0, 2f\n\t"
+				       "\n1:\n\t"
+				       "sub %0, %0, #1\n\t"
+				       "strb wzr, [%1, %0]\n\t"
+				       "cbnz %0, 1b\n\t"
+				       "\n2:\n\t"
+				       : "=&r" ( discard_offset )
+				       : "r" ( dest ), "0" ( len )
+				       : "memory" );
+		return;
+	}
+
+	/* Use "stp" to zero 16 bytes at a time: one initial
+	 * potentially unaligned access, multiple aligned accesses,
+	 * one final potentially unaligned access.
+	 */
+	__asm__ __volatile__ ( "stp xzr, xzr, [%0], #16\n\t"
+			       "bic %0, %0, #15\n\t"
+			       "bic %1, %2, #15\n\t"
+			       "b 2f\n\t"
+			       "\n1:\n\t"
+			       "stp xzr, xzr, [%0], #16\n\t"
+			       "\n2:\n\t"
+			       "cmp %0, %1\n\t"
+			       "bne 1b\n\t"
+			       "stp xzr, xzr, [%2, #-16]\n\t"
+			       : "=&r" ( discard_dest ),
+				 "=&r" ( discard_end )
+			       : "r" ( dest + len ), "0" ( dest )
+			       : "memory", "cc" );
+}
+
+/**
+ * Fill memory region
+ *
+ * @v dest		Destination region
+ * @v len		Length
+ * @v character		Fill character
+ *
+ * The unusual parameter order is to allow for more efficient
+ * tail-calling to arm64_memset() when zeroing a region.
+ */
+void arm64_memset ( void *dest, size_t len, int character ) {
+	size_t discard_offset;
+
+	/* Use optimised zeroing code if applicable */
+	if ( character == 0 ) {
+		arm64_bzero ( dest, len );
+		return;
+	}
+
+	/* Fill one byte at a time.  Calling memset() with a non-zero
+	 * value is relatively rare and unlikely to be
+	 * performance-critical.
+	 */
+	__asm__ __volatile__ ( "cbz %0, 2f\n\t"
+			       "\n1:\n\t"
+			       "sub %0, %0, #1\n\t"
+			       "strb %w2, [%1, %0]\n\t"
+			       "cbnz %0, 1b\n\t"
+			       "\n2:\n\t"
+			       : "=&r" ( discard_offset )
+			       : "r" ( dest ), "r" ( character ), "0" ( len )
+			       : "memory" );
+}
+
+/**
+ * Copy (possibly overlapping) memory region forwards
+ *
+ * @v dest		Destination region
+ * @v src		Source region
+ * @v len		Length
+ */
+void arm64_memmove_forwards ( void *dest, const void *src, size_t len ) {
+	void *discard_dest;
+	const void *discard_src;
+	unsigned long discard_data;
+
+	/* Assume memmove() is not performance-critical, and perform a
+	 * bytewise copy for simplicity.
+	 */
+	__asm__ __volatile__ ( "b 2f\n\t"
+			       "\n1:\n\t"
+			       "ldrb %w2, [%1], #1\n\t"
+			       "strb %w2, [%0], #1\n\t"
+			       "\n2:\n\t"
+			       "cmp %0, %3\n\t"
+			       "bne 1b\n\t"
+			       : "=&r" ( discard_dest ),
+				 "=&r" ( discard_src ),
+				 "=&r" ( discard_data )
+			       : "r" ( dest + len ), "0" ( dest ), "1" ( src )
+			       : "memory" );
+}
+
+/**
+ * Copy (possibly overlapping) memory region backwards
+ *
+ * @v dest		Destination region
+ * @v src		Source region
+ * @v len		Length
+ */
+void arm64_memmove_backwards ( void *dest, const void *src, size_t len ) {
+	size_t discard_offset;
+	unsigned long discard_data;
+
+	/* Assume memmove() is not performance-critical, and perform a
+	 * bytewise copy for simplicity.
+	 */
+	__asm__ __volatile__ ( "cbz %0, 2f\n\t"
+			       "\n1:\n\t"
+			       "sub %0, %0, #1\n\t"
+			       "ldrb %w1, [%3, %0]\n\t"
+			       "strb %w1, [%2, %0]\n\t"
+			       "cbnz %0, 1b\n\t"
+			       "\n2:\n\t"
+			       : "=&r" ( discard_offset ),
+				 "=&r" ( discard_data )
+			       : "r" ( dest ), "r" ( src ), "0" ( len )
+			       : "memory" );
+}
+
+/**
+ * Copy (possibly overlapping) memory region
+ *
+ * @v dest		Destination region
+ * @v src		Source region
+ * @v len		Length
+ */
+void arm64_memmove ( void *dest, const void *src, size_t len ) {
+
+	if ( dest <= src ) {
+		arm64_memmove_forwards ( dest, src, len );
+	} else {
+		arm64_memmove_backwards ( dest, src, len );
+	}
+}
diff --git a/src/arch/arm64/include/bits/string.h b/src/arch/arm64/include/bits/string.h
new file mode 100644
index 00000000..c05fbe34
--- /dev/null
+++ b/src/arch/arm64/include/bits/string.h
@@ -0,0 +1,106 @@
+#ifndef BITS_STRING_H
+#define BITS_STRING_H
+
+FILE_LICENCE ( GPL2_OR_LATER_OR_UBDL );
+
+/** @file
+ *
+ * String functions
+ *
+ */
+
+extern void arm64_bzero ( void *dest, size_t len );
+extern void arm64_memset ( void *dest, size_t len, int character );
+extern void arm64_memcpy ( void *dest, const void *src, size_t len );
+extern void arm64_memmove_forwards ( void *dest, const void *src, size_t len );
+extern void arm64_memmove_backwards ( void *dest, const void *src, size_t len );
+extern void arm64_memmove ( void *dest, const void *src, size_t len );
+
+/**
+ * Fill memory region
+ *
+ * @v dest		Destination region
+ * @v character		Fill character
+ * @v len		Length
+ * @ret dest		Destination region
+ */
+static inline __attribute__ (( always_inline )) void *
+memset ( void *dest, int character, size_t len ) {
+
+	/* Allow gcc to generate inline "stX xzr" instructions for
+	 * small, constant lengths.
+	 */
+	if ( __builtin_constant_p ( character ) && ( character == 0 ) &&
+	     __builtin_constant_p ( len ) && ( len <= 64 ) ) {
+		__builtin_memset ( dest, 0, len );
+		return dest;
+	}
+
+	/* For zeroing larger or non-constant lengths, use the
+	 * optimised variable-length zeroing code.
+	 */
+	if ( __builtin_constant_p ( character ) && ( character == 0 ) ) {
+		arm64_bzero ( dest, len );
+		return dest;
+	}
+
+	/* Not necessarily zeroing: use basic variable-length code */
+	arm64_memset ( dest, len, character );
+	return dest;
+}
+
+/**
+ * Copy memory region
+ *
+ * @v dest		Destination region
+ * @v src		Source region
+ * @v len		Length
+ * @ret dest		Destination region
+ */
+static inline __attribute__ (( always_inline )) void *
+memcpy ( void *dest, const void *src, size_t len ) {
+
+	/* Allow gcc to generate inline "ldX"/"stX" instructions for
+	 * small, constant lengths.
+	 */
+	if ( __builtin_constant_p ( len ) && ( len <= 64 ) ) {
+		__builtin_memcpy ( dest, src, len );
+		return dest;
+	}
+
+	/* Otherwise, use variable-length code */
+	arm64_memcpy ( dest, src, len );
+	return dest;
+}
+
+/**
+ * Copy (possibly overlapping) memory region
+ *
+ * @v dest		Destination region
+ * @v src		Source region
+ * @v len		Length
+ * @ret dest		Destination region
+ */
+static inline __attribute__ (( always_inline )) void *
+memmove ( void *dest, const void *src, size_t len ) {
+	ssize_t offset = ( dest - src );
+
+	/* If required direction of copy is known at build time, then
+	 * use the appropriate forwards/backwards copy directly.
+	 */
+	if ( __builtin_constant_p ( offset ) ) {
+		if ( offset <= 0 ) {
+			arm64_memmove_forwards ( dest, src, len );
+			return dest;
+		} else {
+			arm64_memmove_backwards ( dest, src, len );
+			return dest;
+		}
+	}
+
+	/* Otherwise, use ambidirectional copy */
+	arm64_memmove ( dest, src, len );
+	return dest;
+}
+
+#endif /* BITS_STRING_H */