Fixed MTP to work with TWRP

arch/alpha/lib/Makefile

@@ -0,0 +1,56 @@
+#
+# Makefile for alpha-specific library files..
+#
+
+asflags-y := $(KBUILD_CFLAGS)
+ccflags-y := -Werror
+
+# Many of these routines have implementations tuned for ev6.
+# Choose them iff we're targeting ev6 specifically.
+ev6-$(CONFIG_ALPHA_EV6) := ev6-
+
+# Several make use of the cttz instruction introduced in ev67.
+ev67-$(CONFIG_ALPHA_EV67) := ev67-
+
+lib-y =	__divqu.o __remqu.o __divlu.o __remlu.o \
+	udelay.o \
+	$(ev6-y)memset.o \
+	$(ev6-y)memcpy.o \
+	memmove.o \
+	checksum.o \
+	csum_partial_copy.o \
+	$(ev67-y)strlen.o \
+	$(ev67-y)strcat.o \
+	strcpy.o \
+	$(ev67-y)strncat.o \
+	strncpy.o \
+	$(ev6-y)stxcpy.o \
+	$(ev6-y)stxncpy.o \
+	$(ev67-y)strchr.o \
+	$(ev67-y)strrchr.o \
+	$(ev6-y)memchr.o \
+	$(ev6-y)copy_user.o \
+	$(ev6-y)clear_user.o \
+	$(ev6-y)csum_ipv6_magic.o \
+	$(ev6-y)clear_page.o \
+	$(ev6-y)copy_page.o \
+	fpreg.o \
+	callback_srm.o srm_puts.o srm_printk.o \
+	fls.o
+
+lib-$(CONFIG_SMP) += dec_and_lock.o
+
+# The division routines are built from single source, with different defines.
+AFLAGS___divqu.o = -DDIV
+AFLAGS___remqu.o =       -DREM
+AFLAGS___divlu.o = -DDIV       -DINTSIZE
+AFLAGS___remlu.o =       -DREM -DINTSIZE
+
+$(obj)/__divqu.o: $(obj)/$(ev6-y)divide.S
+	$(cmd_as_o_S)
+$(obj)/__remqu.o: $(obj)/$(ev6-y)divide.S
+	$(cmd_as_o_S)
+$(obj)/__divlu.o: $(obj)/$(ev6-y)divide.S
+	$(cmd_as_o_S)
+$(obj)/__remlu.o: $(obj)/$(ev6-y)divide.S
+	$(cmd_as_o_S)

arch/alpha/lib/callback_srm.S

@@ -0,0 +1,103 @@
+/*
+ *	arch/alpha/lib/callback_srm.S
+ */
+
+#include <asm/console.h>
+
+.text
+#define HWRPB_CRB_OFFSET 0xc0
+
+#if defined(CONFIG_ALPHA_SRM) || defined(CONFIG_ALPHA_GENERIC)
+.align 4
+srm_dispatch:
+#if defined(CONFIG_ALPHA_GENERIC)
+	ldl	$4,alpha_using_srm
+	beq	$4,nosrm
+#endif
+	ldq	$0,hwrpb	# gp is set up by CALLBACK macro.
+	ldl	$25,0($25)	# Pick up the wrapper data.
+	mov	$20,$21		# Shift arguments right.
+	mov	$19,$20
+	ldq	$1,HWRPB_CRB_OFFSET($0)
+	mov	$18,$19
+	mov	$17,$18
+	mov	$16,$17
+	addq	$0,$1,$2	# CRB address
+	ldq	$27,0($2)	# DISPATCH procedure descriptor (VMS call std)
+	extwl	$25,0,$16	# SRM callback function code
+	ldq	$3,8($27)	# call address
+	extwl	$25,2,$25	# argument information (VMS calling std)
+	jmp	($3)		# Return directly to caller of wrapper.
+
+.align 4
+.globl	srm_fixup
+.ent	srm_fixup
+srm_fixup:
+	ldgp	$29,0($27)
+#if defined(CONFIG_ALPHA_GENERIC)
+	ldl	$4,alpha_using_srm
+	beq	$4,nosrm
+#endif
+	ldq	$0,hwrpb
+	ldq	$1,HWRPB_CRB_OFFSET($0)
+	addq	$0,$1,$2	# CRB address
+	ldq	$27,16($2)	# VA of FIXUP procedure descriptor
+	ldq	$3,8($27)	# call address
+	lda	$25,2($31)	# two integer arguments
+	jmp	($3)		# Return directly to caller of srm_fixup.
+.end    srm_fixup
+
+#if defined(CONFIG_ALPHA_GENERIC)
+.align 3
+nosrm:
+	lda	$0,-1($31)
+	ret
+#endif
+
+#define CALLBACK(NAME, CODE, ARG_CNT) \
+.align 4; .globl callback_##NAME; .ent callback_##NAME; callback_##NAME##: \
+ldgp $29,0($27); br $25,srm_dispatch; .word CODE, ARG_CNT; .end callback_##NAME
+
+#else /* defined(CONFIG_ALPHA_SRM) || defined(CONFIG_ALPHA_GENERIC) */
+
+#define CALLBACK(NAME, CODE, ARG_CNT) \
+.align 3; .globl callback_##NAME; .ent callback_##NAME; callback_##NAME##: \
+lda $0,-1($31); ret; .end callback_##NAME
+
+.align 3
+.globl	srm_fixup
+.ent	srm_fixup
+srm_fixup:
+	lda	$0,-1($31)
+	ret
+.end	srm_fixup
+#endif /* defined(CONFIG_ALPHA_SRM) || defined(CONFIG_ALPHA_GENERIC) */
+
+CALLBACK(puts, CCB_PUTS, 4)
+CALLBACK(open, CCB_OPEN, 3)
+CALLBACK(close, CCB_CLOSE, 2)
+CALLBACK(read, CCB_READ, 5)
+CALLBACK(open_console, CCB_OPEN_CONSOLE, 1)
+CALLBACK(close_console, CCB_CLOSE_CONSOLE, 1)
+CALLBACK(getenv, CCB_GET_ENV, 4)
+CALLBACK(setenv, CCB_SET_ENV, 4)
+CALLBACK(getc, CCB_GETC, 2)
+CALLBACK(reset_term, CCB_RESET_TERM, 2)
+CALLBACK(term_int, CCB_SET_TERM_INT, 3)
+CALLBACK(term_ctl, CCB_SET_TERM_CTL, 3)
+CALLBACK(process_keycode, CCB_PROCESS_KEYCODE, 3)
+CALLBACK(ioctl, CCB_IOCTL, 6)
+CALLBACK(write, CCB_WRITE, 5)
+CALLBACK(reset_env, CCB_RESET_ENV, 4)
+CALLBACK(save_env, CCB_SAVE_ENV, 1)
+CALLBACK(pswitch, CCB_PSWITCH, 3)
+CALLBACK(bios_emul, CCB_BIOS_EMUL, 5)
+	
+.data
+__alpha_using_srm:		# For use by bootpheader
+	.long 7			# value is not 1 for link debugging
+	.weak alpha_using_srm; alpha_using_srm = __alpha_using_srm
+__callback_init_done:		# For use by bootpheader
+	.long 7			# value is not 1 for link debugging
+	.weak callback_init_done; callback_init_done = __callback_init_done
+

arch/alpha/lib/checksum.c

@@ -0,0 +1,184 @@
+/*
+ * arch/alpha/lib/checksum.c
+ *
+ * This file contains network checksum routines that are better done
+ * in an architecture-specific manner due to speed..
+ * Comments in other versions indicate that the algorithms are from RFC1071
+ *
+ * accelerated versions (and 21264 assembly versions ) contributed by
+ *	Rick Gorton	<rick.gorton@alpha-processor.com>
+ */
+ 
+#include <linux/module.h>
+#include <linux/string.h>
+
+#include <asm/byteorder.h>
+
+static inline unsigned short from64to16(unsigned long x)
+{
+	/* Using extract instructions is a bit more efficient
+	   than the original shift/bitmask version.  */
+
+	union {
+		unsigned long	ul;
+		unsigned int	ui[2];
+		unsigned short	us[4];
+	} in_v, tmp_v, out_v;
+
+	in_v.ul = x;
+	tmp_v.ul = (unsigned long) in_v.ui[0] + (unsigned long) in_v.ui[1];
+
+	/* Since the bits of tmp_v.sh[3] are going to always be zero,
+	   we don't have to bother to add that in.  */
+	out_v.ul = (unsigned long) tmp_v.us[0] + (unsigned long) tmp_v.us[1]
+			+ (unsigned long) tmp_v.us[2];
+
+	/* Similarly, out_v.us[2] is always zero for the final add.  */
+	return out_v.us[0] + out_v.us[1];
+}
+
+/*
+ * computes the checksum of the TCP/UDP pseudo-header
+ * returns a 16-bit checksum, already complemented.
+ */
+__sum16 csum_tcpudp_magic(__be32 saddr, __be32 daddr,
+				   unsigned short len,
+				   unsigned short proto,
+				   __wsum sum)
+{
+	return (__force __sum16)~from64to16(
+		(__force u64)saddr + (__force u64)daddr +
+		(__force u64)sum + ((len + proto) << 8));
+}
+
+__wsum csum_tcpudp_nofold(__be32 saddr, __be32 daddr,
+				   unsigned short len,
+				   unsigned short proto,
+				   __wsum sum)
+{
+	unsigned long result;
+
+	result = (__force u64)saddr + (__force u64)daddr +
+		 (__force u64)sum + ((len + proto) << 8);
+
+	/* Fold down to 32-bits so we don't lose in the typedef-less 
+	   network stack.  */
+	/* 64 to 33 */
+	result = (result & 0xffffffff) + (result >> 32);
+	/* 33 to 32 */
+	result = (result & 0xffffffff) + (result >> 32);
+	return (__force __wsum)result;
+}
+EXPORT_SYMBOL(csum_tcpudp_nofold);
+
+/*
+ * Do a 64-bit checksum on an arbitrary memory area..
+ *
+ * This isn't a great routine, but it's not _horrible_ either. The
+ * inner loop could be unrolled a bit further, and there are better
+ * ways to do the carry, but this is reasonable.
+ */
+static inline unsigned long do_csum(const unsigned char * buff, int len)
+{
+	int odd, count;
+	unsigned long result = 0;
+
+	if (len <= 0)
+		goto out;
+	odd = 1 & (unsigned long) buff;
+	if (odd) {
+		result = *buff << 8;
+		len--;
+		buff++;
+	}
+	count = len >> 1;		/* nr of 16-bit words.. */
+	if (count) {
+		if (2 & (unsigned long) buff) {
+			result += *(unsigned short *) buff;
+			count--;
+			len -= 2;
+			buff += 2;
+		}
+		count >>= 1;		/* nr of 32-bit words.. */
+		if (count) {
+			if (4 & (unsigned long) buff) {
+				result += *(unsigned int *) buff;
+				count--;
+				len -= 4;
+				buff += 4;
+			}
+			count >>= 1;	/* nr of 64-bit words.. */
+			if (count) {
+				unsigned long carry = 0;
+				do {
+					unsigned long w = *(unsigned long *) buff;
+					count--;
+					buff += 8;
+					result += carry;
+					result += w;
+					carry = (w > result);
+				} while (count);
+				result += carry;
+				result = (result & 0xffffffff) + (result >> 32);
+			}
+			if (len & 4) {
+				result += *(unsigned int *) buff;
+				buff += 4;
+			}
+		}
+		if (len & 2) {
+			result += *(unsigned short *) buff;
+			buff += 2;
+		}
+	}
+	if (len & 1)
+		result += *buff;
+	result = from64to16(result);
+	if (odd)
+		result = ((result >> 8) & 0xff) | ((result & 0xff) << 8);
+out:
+	return result;
+}
+
+/*
+ *	This is a version of ip_compute_csum() optimized for IP headers,
+ *	which always checksum on 4 octet boundaries.
+ */
+__sum16 ip_fast_csum(const void *iph, unsigned int ihl)
+{
+	return (__force __sum16)~do_csum(iph,ihl*4);
+}
+
+/*
+ * computes the checksum of a memory block at buff, length len,
+ * and adds in "sum" (32-bit)
+ *
+ * returns a 32-bit number suitable for feeding into itself
+ * or csum_tcpudp_magic
+ *
+ * this function must be called with even lengths, except
+ * for the last fragment, which may be odd
+ *
+ * it's best to have buff aligned on a 32-bit boundary
+ */
+__wsum csum_partial(const void *buff, int len, __wsum sum)
+{
+	unsigned long result = do_csum(buff, len);
+
+	/* add in old sum, and carry.. */
+	result += (__force u32)sum;
+	/* 32+c bits -> 32 bits */
+	result = (result & 0xffffffff) + (result >> 32);
+	return (__force __wsum)result;
+}
+
+EXPORT_SYMBOL(csum_partial);
+
+/*
+ * this routine is used for miscellaneous IP-like checksums, mainly
+ * in icmp.c
+ */
+__sum16 ip_compute_csum(const void *buff, int len)
+{
+	return (__force __sum16)~from64to16(do_csum(buff,len));
+}

arch/alpha/lib/clear_page.S

@@ -0,0 +1,39 @@
+/*
+ * arch/alpha/lib/clear_page.S
+ *
+ * Zero an entire page.
+ */
+
+	.text
+	.align 4
+	.global clear_page
+	.ent clear_page
+clear_page:
+	.prologue 0
+
+	lda	$0,128
+	nop
+	unop
+	nop
+
+1:	stq	$31,0($16)
+	stq	$31,8($16)
+	stq	$31,16($16)
+	stq	$31,24($16)
+
+	stq	$31,32($16)
+	stq	$31,40($16)
+	stq	$31,48($16)
+	subq	$0,1,$0
+
+	stq	$31,56($16)
+	addq	$16,64,$16
+	unop
+	bne	$0,1b
+
+	ret
+	nop
+	unop
+	nop
+
+	.end clear_page

arch/alpha/lib/clear_user.S

@@ -0,0 +1,113 @@
+/*
+ * arch/alpha/lib/clear_user.S
+ * Contributed by Richard Henderson <rth@tamu.edu>
+ *
+ * Zero user space, handling exceptions as we go.
+ *
+ * We have to make sure that $0 is always up-to-date and contains the
+ * right "bytes left to zero" value (and that it is updated only _after_
+ * a successful copy).  There is also some rather minor exception setup
+ * stuff.
+ *
+ * NOTE! This is not directly C-callable, because the calling semantics
+ * are different:
+ *
+ * Inputs:
+ *	length in $0
+ *	destination address in $6
+ *	exception pointer in $7
+ *	return address in $28 (exceptions expect it there)
+ *
+ * Outputs:
+ *	bytes left to copy in $0
+ *
+ * Clobbers:
+ *	$1,$2,$3,$4,$5,$6
+ */
+
+/* Allow an exception for an insn; exit if we get one.  */
+#define EX(x,y...)			\
+	99: x,##y;			\
+	.section __ex_table,"a";	\
+	.long 99b - .;			\
+	lda $31, $exception-99b($31); 	\
+	.previous
+
+	.set noat
+	.set noreorder
+	.align 4
+
+	.globl __do_clear_user
+	.ent __do_clear_user
+	.frame	$30, 0, $28
+	.prologue 0
+
+$loop:
+	and	$1, 3, $4	# e0    :
+	beq	$4, 1f		# .. e1 :
+
+0:	EX( stq_u $31, 0($6) )	# e0    : zero one word
+	subq	$0, 8, $0	# .. e1 :
+	subq	$4, 1, $4	# e0    :
+	addq	$6, 8, $6	# .. e1 :
+	bne	$4, 0b		# e1    :
+	unop			#       :
+
+1:	bic	$1, 3, $1	# e0    :
+	beq	$1, $tail	# .. e1 :
+
+2:	EX( stq_u $31, 0($6) )	# e0    : zero four words
+	subq	$0, 8, $0	# .. e1 :
+	EX( stq_u $31, 8($6) )	# e0    :
+	subq	$0, 8, $0	# .. e1 :
+	EX( stq_u $31, 16($6) )	# e0    :
+	subq	$0, 8, $0	# .. e1 :
+	EX( stq_u $31, 24($6) )	# e0    :
+	subq	$0, 8, $0	# .. e1 :
+	subq	$1, 4, $1	# e0    :
+	addq	$6, 32, $6	# .. e1 :
+	bne	$1, 2b		# e1    :
+
+$tail:
+	bne	$2, 1f		# e1    : is there a tail to do?
+	ret	$31, ($28), 1	# .. e1 :
+
+1:	EX( ldq_u $5, 0($6) )	# e0    :
+	clr	$0		# .. e1 :
+	nop			# e1    :
+	mskqh	$5, $0, $5	# e0    :
+	EX( stq_u $5, 0($6) )	# e0    :
+	ret	$31, ($28), 1	# .. e1 :
+
+__do_clear_user:
+	and	$6, 7, $4	# e0    : find dest misalignment
+	beq	$0, $zerolength # .. e1 :
+	addq	$0, $4, $1	# e0    : bias counter
+	and	$1, 7, $2	# e1    : number of bytes in tail
+	srl	$1, 3, $1	# e0    :
+	beq	$4, $loop	# .. e1 :
+
+	EX( ldq_u $5, 0($6) )	# e0    : load dst word to mask back in
+	beq	$1, $oneword	# .. e1 : sub-word store?
+
+	mskql	$5, $6, $5	# e0    : take care of misaligned head
+	addq	$6, 8, $6	# .. e1 :
+	EX( stq_u $5, -8($6) )	# e0    :
+	addq	$0, $4, $0	# .. e1 : bytes left -= 8 - misalignment
+	subq	$1, 1, $1	# e0    :
+	subq	$0, 8, $0	# .. e1 :
+	br	$loop		# e1    :
+	unop			#       :
+
+$oneword:
+	mskql	$5, $6, $4	# e0    :
+	mskqh	$5, $2, $5	# e0    :
+	or	$5, $4, $5	# e1    :
+	EX( stq_u $5, 0($6) )	# e0    :
+	clr	$0		# .. e1 :
+
+$zerolength:
+$exception:
+	ret	$31, ($28), 1	# .. e1 :
+
+	.end __do_clear_user

arch/alpha/lib/copy_page.S

@@ -0,0 +1,49 @@
+/*
+ * arch/alpha/lib/copy_page.S
+ *
+ * Copy an entire page.
+ */
+
+	.text
+	.align 4
+	.global copy_page
+	.ent copy_page
+copy_page:
+	.prologue 0
+
+	lda	$18,128
+	nop
+	unop
+	nop
+
+1:	ldq	$0,0($17)
+	ldq	$1,8($17)
+	ldq	$2,16($17)
+	ldq	$3,24($17)
+
+	ldq	$4,32($17)
+	ldq	$5,40($17)
+	ldq	$6,48($17)
+	ldq	$7,56($17)
+
+	stq	$0,0($16)
+	subq	$18,1,$18
+	stq	$1,8($16)
+	addq	$17,64,$17
+
+	stq	$2,16($16)
+	stq	$3,24($16)
+	stq	$4,32($16)
+	stq	$5,40($16)
+
+	stq	$6,48($16)
+	stq	$7,56($16)
+	addq	$16,64,$16
+	bne	$18, 1b
+
+	ret
+	nop
+	unop
+	nop
+
+	.end copy_page

arch/alpha/lib/copy_user.S

@@ -0,0 +1,145 @@
+/*
+ * arch/alpha/lib/copy_user.S
+ *
+ * Copy to/from user space, handling exceptions as we go..  This
+ * isn't exactly pretty.
+ *
+ * This is essentially the same as "memcpy()", but with a few twists.
+ * Notably, we have to make sure that $0 is always up-to-date and
+ * contains the right "bytes left to copy" value (and that it is updated
+ * only _after_ a successful copy). There is also some rather minor
+ * exception setup stuff..
+ *
+ * NOTE! This is not directly C-callable, because the calling semantics are
+ * different:
+ *
+ * Inputs:
+ *	length in $0
+ *	destination address in $6
+ *	source address in $7
+ *	return address in $28
+ *
+ * Outputs:
+ *	bytes left to copy in $0
+ *
+ * Clobbers:
+ *	$1,$2,$3,$4,$5,$6,$7
+ */
+
+/* Allow an exception for an insn; exit if we get one.  */
+#define EXI(x,y...)			\
+	99: x,##y;			\
+	.section __ex_table,"a";	\
+	.long 99b - .;			\
+	lda $31, $exitin-99b($31);	\
+	.previous
+
+#define EXO(x,y...)			\
+	99: x,##y;			\
+	.section __ex_table,"a";	\
+	.long 99b - .;			\
+	lda $31, $exitout-99b($31);	\
+	.previous
+
+	.set noat
+	.align 4
+	.globl __copy_user
+	.ent __copy_user
+__copy_user:
+	.prologue 0
+	and $6,7,$3
+	beq $0,$35
+	beq $3,$36
+	subq $3,8,$3
+	.align 4
+$37:
+	EXI( ldq_u $1,0($7) )
+	EXO( ldq_u $2,0($6) )
+	extbl $1,$7,$1
+	mskbl $2,$6,$2
+	insbl $1,$6,$1
+	addq $3,1,$3
+	bis $1,$2,$1
+	EXO( stq_u $1,0($6) )
+	subq $0,1,$0
+	addq $6,1,$6
+	addq $7,1,$7
+	beq $0,$41
+	bne $3,$37
+$36:
+	and $7,7,$1
+	bic $0,7,$4
+	beq $1,$43
+	beq $4,$48
+	EXI( ldq_u $3,0($7) )
+	.align 4
+$50:
+	EXI( ldq_u $2,8($7) )
+	subq $4,8,$4
+	extql $3,$7,$3
+	extqh $2,$7,$1
+	bis $3,$1,$1
+	EXO( stq $1,0($6) )
+	addq $7,8,$7
+	subq $0,8,$0
+	addq $6,8,$6
+	bis $2,$2,$3
+	bne $4,$50
+$48:
+	beq $0,$41
+	.align 4
+$57:
+	EXI( ldq_u $1,0($7) )
+	EXO( ldq_u $2,0($6) )
+	extbl $1,$7,$1
+	mskbl $2,$6,$2
+	insbl $1,$6,$1
+	bis $1,$2,$1
+	EXO( stq_u $1,0($6) )
+	subq $0,1,$0
+	addq $6,1,$6
+	addq $7,1,$7
+	bne $0,$57
+	br $31,$41
+	.align 4
+$43:
+	beq $4,$65
+	.align 4
+$66:
+	EXI( ldq $1,0($7) )
+	subq $4,8,$4
+	EXO( stq $1,0($6) )
+	addq $7,8,$7
+	subq $0,8,$0
+	addq $6,8,$6
+	bne $4,$66
+$65:
+	beq $0,$41
+	EXI( ldq $2,0($7) )
+	EXO( ldq $1,0($6) )
+	mskql $2,$0,$2
+	mskqh $1,$0,$1
+	bis $2,$1,$2
+	EXO( stq $2,0($6) )
+	bis $31,$31,$0
+$41:
+$35:
+$exitout:
+	ret $31,($28),1
+
+$exitin:
+	/* A stupid byte-by-byte zeroing of the rest of the output
+	   buffer.  This cures security holes by never leaving 
+	   random kernel data around to be copied elsewhere.  */
+
+	mov $0,$1
+$101:
+	EXO ( ldq_u $2,0($6) )
+	subq $1,1,$1
+	mskbl $2,$6,$2
+	EXO ( stq_u $2,0($6) )
+	addq $6,1,$6
+	bgt $1,$101
+	ret $31,($28),1
+
+	.end __copy_user

arch/alpha/lib/csum_ipv6_magic.S

@@ -0,0 +1,115 @@
+/*
+ * arch/alpha/lib/csum_ipv6_magic.S
+ * Contributed by Richard Henderson <rth@tamu.edu>
+ *
+ * unsigned short csum_ipv6_magic(struct in6_addr *saddr,
+ *                                struct in6_addr *daddr,
+ *                                __u32 len,
+ *                                unsigned short proto,
+ *                                unsigned int csum);
+ *
+ * Misalignment handling (which costs 16 instructions / 8 cycles)
+ * added by Ivan Kokshaysky <ink@jurassic.park.msu.ru>
+ */
+
+	.globl csum_ipv6_magic
+	.align 4
+	.ent csum_ipv6_magic
+	.frame $30,0,$26,0
+csum_ipv6_magic:
+	.prologue 0
+
+	ldq_u	$0,0($16)	# e0    : load src & dst addr words
+	zapnot	$20,15,$20	# .. e1 : zero extend incoming csum
+	extqh	$18,1,$4	# e0    : byte swap len & proto while we wait
+	ldq_u	$21,7($16)	# .. e1 : handle misalignment
+
+	extbl	$18,1,$5	# e0	:
+	ldq_u	$1,8($16)	# .. e1 :
+	extbl	$18,2,$6	# e0 	:
+	ldq_u	$22,15($16)	# .. e1 :
+
+	extbl	$18,3,$18	# e0	:
+	ldq_u	$2,0($17)	# .. e1 :
+	sra	$4,32,$4	# e0	:
+	ldq_u	$23,7($17)	# .. e1 :
+
+	extql	$0,$16,$0	# e0	:
+	ldq_u	$3,8($17)	# .. e1 :
+	extqh	$21,$16,$21	# e0	:
+	ldq_u	$24,15($17)	# .. e1 :
+
+	sll	$5,16,$5	# e0	:
+	or	$0,$21,$0	# .. e1 : 1st src word complete
+	extql	$1,$16,$1	# e0	:
+	addq	$20,$0,$20	# .. e1 : begin summing the words
+
+	extqh	$22,$16,$22	# e0	:
+	cmpult	$20,$0,$0	# .. e1 :
+	sll	$6,8,$6		# e0	:
+	or	$1,$22,$1	# .. e1 : 2nd src word complete
+
+	extql	$2,$17,$2	# e0	:
+	or	$4,$18,$18	# .. e1 :
+	extqh	$23,$17,$23	# e0	:
+	or	$5,$6,$5	# .. e1 :
+
+	extql	$3,$17,$3	# e0	:
+	or	$2,$23,$2	# .. e1 : 1st dst word complete
+	extqh	$24,$17,$24	# e0	:
+	or	$18,$5,$18	# .. e1 : len complete
+
+	extwh	$19,7,$7	# e0    :
+	or	$3,$24,$3	# .. e1 : 2nd dst word complete
+	extbl	$19,1,$19	# e0    :
+	addq	$20,$1,$20	# .. e1 :
+
+	or	$19,$7,$19	# e0    :
+	cmpult	$20,$1,$1	# .. e1 :
+	sll	$19,48,$19	# e0    :
+	nop			# .. e0 :
+
+	sra	$19,32,$19	# e0    : proto complete
+	addq	$20,$2,$20	# .. e1 :
+	cmpult	$20,$2,$2	# e0    :
+	addq	$20,$3,$20	# .. e1 :
+
+	cmpult	$20,$3,$3	# e0    :
+	addq	$20,$18,$20	# .. e1 :
+	cmpult	$20,$18,$18	# e0    :
+	addq	$20,$19,$20	# .. e1 :
+
+	cmpult	$20,$19,$19	# e0    :
+	addq	$0,$1,$0	# .. e1 : merge the carries back into the csum
+	addq	$2,$3,$2	# e0    :
+	addq	$18,$19,$18	# .. e1 :
+
+	addq	$0,$2,$0	# e0    :
+	addq	$20,$18,$20	# .. e1 :
+	addq	$0,$20,$0	# e0    :
+	unop			#       :
+
+	extwl	$0,2,$2		# e0    : begin folding the 64-bit value
+	zapnot	$0,3,$3		# .. e1 :
+	extwl	$0,4,$1		# e0    :
+	addq	$2,$3,$3	# .. e1 :
+
+	extwl	$0,6,$0		# e0    :
+	addq	$3,$1,$3	# .. e1 :
+	addq	$0,$3,$0	# e0    :
+	unop			#       :
+
+	extwl	$0,2,$1		# e0    : fold 18-bit value
+	zapnot	$0,3,$0		# .. e1 :
+	addq	$0,$1,$0	# e0    :
+	unop			#       :
+
+	extwl	$0,2,$1		# e0    : fold 17-bit value
+	zapnot	$0,3,$0		# .. e1 :
+	addq	$0,$1,$0	# e0    :
+	not	$0,$0		# .. e1 : and complement.
+
+	zapnot	$0,3,$0		# e0    :
+	ret			# .. e1 :
+
+	.end csum_ipv6_magic

arch/alpha/lib/csum_partial_copy.c

@@ -0,0 +1,388 @@
+/*
+ * csum_partial_copy - do IP checksumming and copy
+ *
+ * (C) Copyright 1996 Linus Torvalds
+ * accelerated versions (and 21264 assembly versions ) contributed by
+ *	Rick Gorton	<rick.gorton@alpha-processor.com>
+ *
+ * Don't look at this too closely - you'll go mad. The things
+ * we do for performance..
+ */
+
+#include <linux/types.h>
+#include <linux/string.h>
+#include <asm/uaccess.h>
+
+
+#define ldq_u(x,y) \
+__asm__ __volatile__("ldq_u %0,%1":"=r" (x):"m" (*(const unsigned long *)(y)))
+
+#define stq_u(x,y) \
+__asm__ __volatile__("stq_u %1,%0":"=m" (*(unsigned long *)(y)):"r" (x))
+
+#define extql(x,y,z) \
+__asm__ __volatile__("extql %1,%2,%0":"=r" (z):"r" (x),"r" (y))
+
+#define extqh(x,y,z) \
+__asm__ __volatile__("extqh %1,%2,%0":"=r" (z):"r" (x),"r" (y))
+
+#define mskql(x,y,z) \
+__asm__ __volatile__("mskql %1,%2,%0":"=r" (z):"r" (x),"r" (y))
+
+#define mskqh(x,y,z) \
+__asm__ __volatile__("mskqh %1,%2,%0":"=r" (z):"r" (x),"r" (y))
+
+#define insql(x,y,z) \
+__asm__ __volatile__("insql %1,%2,%0":"=r" (z):"r" (x),"r" (y))
+
+#define insqh(x,y,z) \
+__asm__ __volatile__("insqh %1,%2,%0":"=r" (z):"r" (x),"r" (y))
+
+
+#define __get_user_u(x,ptr)				\
+({							\
+	long __guu_err;					\
+	__asm__ __volatile__(				\
+	"1:	ldq_u %0,%2\n"				\
+	"2:\n"						\
+	".section __ex_table,\"a\"\n"			\
+	"	.long 1b - .\n"				\
+	"	lda %0,2b-1b(%1)\n"			\
+	".previous"					\
+		: "=r"(x), "=r"(__guu_err)		\
+		: "m"(__m(ptr)), "1"(0));		\
+	__guu_err;					\
+})
+
+#define __put_user_u(x,ptr)				\
+({							\
+	long __puu_err;					\
+	__asm__ __volatile__(				\
+	"1:	stq_u %2,%1\n"				\
+	"2:\n"						\
+	".section __ex_table,\"a\"\n"			\
+	"	.long 1b - ."				\
+	"	lda $31,2b-1b(%0)\n"			\
+	".previous"					\
+		: "=r"(__puu_err)			\
+		: "m"(__m(addr)), "rJ"(x), "0"(0));	\
+	__puu_err;					\
+})
+
+
+static inline unsigned short from64to16(unsigned long x)
+{
+	/* Using extract instructions is a bit more efficient
+	   than the original shift/bitmask version.  */
+
+	union {
+		unsigned long	ul;
+		unsigned int	ui[2];
+		unsigned short	us[4];
+	} in_v, tmp_v, out_v;
+
+	in_v.ul = x;
+	tmp_v.ul = (unsigned long) in_v.ui[0] + (unsigned long) in_v.ui[1];
+
+	/* Since the bits of tmp_v.sh[3] are going to always be zero,
+	   we don't have to bother to add that in.  */
+	out_v.ul = (unsigned long) tmp_v.us[0] + (unsigned long) tmp_v.us[1]
+			+ (unsigned long) tmp_v.us[2];
+
+	/* Similarly, out_v.us[2] is always zero for the final add.  */
+	return out_v.us[0] + out_v.us[1];
+}
+
+
+
+/*
+ * Ok. This isn't fun, but this is the EASY case.
+ */
+static inline unsigned long
+csum_partial_cfu_aligned(const unsigned long __user *src, unsigned long *dst,
+			 long len, unsigned long checksum,
+			 int *errp)
+{
+	unsigned long carry = 0;
+	int err = 0;
+
+	while (len >= 0) {
+		unsigned long word;
+		err |= __get_user(word, src);
+		checksum += carry;
+		src++;
+		checksum += word;
+		len -= 8;
+		carry = checksum < word;
+		*dst = word;
+		dst++;
+	}
+	len += 8;
+	checksum += carry;
+	if (len) {
+		unsigned long word, tmp;
+		err |= __get_user(word, src);
+		tmp = *dst;
+		mskql(word, len, word);
+		checksum += word;
+		mskqh(tmp, len, tmp);
+		carry = checksum < word;
+		*dst = word | tmp;
+		checksum += carry;
+	}
+	if (err && errp) *errp = err;
+	return checksum;
+}
+
+/*
+ * This is even less fun, but this is still reasonably
+ * easy.
+ */
+static inline unsigned long
+csum_partial_cfu_dest_aligned(const unsigned long __user *src,
+			      unsigned long *dst,
+			      unsigned long soff,
+			      long len, unsigned long checksum,
+			      int *errp)
+{
+	unsigned long first;
+	unsigned long word, carry;
+	unsigned long lastsrc = 7+len+(unsigned long)src;
+	int err = 0;
+
+	err |= __get_user_u(first,src);
+	carry = 0;
+	while (len >= 0) {
+		unsigned long second;
+
+		err |= __get_user_u(second, src+1);
+		extql(first, soff, word);
+		len -= 8;
+		src++;
+		extqh(second, soff, first);
+		checksum += carry;
+		word |= first;
+		first = second;
+		checksum += word;
+		*dst = word;
+		dst++;
+		carry = checksum < word;
+	}
+	len += 8;
+	checksum += carry;
+	if (len) {
+		unsigned long tmp;
+		unsigned long second;
+		err |= __get_user_u(second, lastsrc);
+		tmp = *dst;
+		extql(first, soff, word);
+		extqh(second, soff, first);
+		word |= first;
+		mskql(word, len, word);
+		checksum += word;
+		mskqh(tmp, len, tmp);
+		carry = checksum < word;
+		*dst = word | tmp;
+		checksum += carry;
+	}
+	if (err && errp) *errp = err;
+	return checksum;
+}
+
+/*
+ * This is slightly less fun than the above..
+ */
+static inline unsigned long
+csum_partial_cfu_src_aligned(const unsigned long __user *src,
+			     unsigned long *dst,
+			     unsigned long doff,
+			     long len, unsigned long checksum,
+			     unsigned long partial_dest,
+			     int *errp)
+{
+	unsigned long carry = 0;
+	unsigned long word;
+	unsigned long second_dest;
+	int err = 0;
+
+	mskql(partial_dest, doff, partial_dest);
+	while (len >= 0) {
+		err |= __get_user(word, src);
+		len -= 8;
+		insql(word, doff, second_dest);
+		checksum += carry;
+		stq_u(partial_dest | second_dest, dst);
+		src++;
+		checksum += word;
+		insqh(word, doff, partial_dest);
+		carry = checksum < word;
+		dst++;
+	}
+	len += 8;
+	if (len) {
+		checksum += carry;
+		err |= __get_user(word, src);
+		mskql(word, len, word);
+		len -= 8;
+		checksum += word;
+		insql(word, doff, second_dest);
+		len += doff;
+		carry = checksum < word;
+		partial_dest |= second_dest;
+		if (len >= 0) {
+			stq_u(partial_dest, dst);
+			if (!len) goto out;
+			dst++;
+			insqh(word, doff, partial_dest);
+		}
+		doff = len;
+	}
+	ldq_u(second_dest, dst);
+	mskqh(second_dest, doff, second_dest);
+	stq_u(partial_dest | second_dest, dst);
+out:
+	checksum += carry;
+	if (err && errp) *errp = err;
+	return checksum;
+}
+
+/*
+ * This is so totally un-fun that it's frightening. Don't
+ * look at this too closely, you'll go blind.
+ */
+static inline unsigned long
+csum_partial_cfu_unaligned(const unsigned long __user * src,
+			   unsigned long * dst,
+			   unsigned long soff, unsigned long doff,
+			   long len, unsigned long checksum,
+			   unsigned long partial_dest,
+			   int *errp)
+{
+	unsigned long carry = 0;
+	unsigned long first;
+	unsigned long lastsrc;
+	int err = 0;
+
+	err |= __get_user_u(first, src);
+	lastsrc = 7+len+(unsigned long)src;
+	mskql(partial_dest, doff, partial_dest);
+	while (len >= 0) {
+		unsigned long second, word;
+		unsigned long second_dest;
+
+		err |= __get_user_u(second, src+1);
+		extql(first, soff, word);
+		checksum += carry;
+		len -= 8;
+		extqh(second, soff, first);
+		src++;
+		word |= first;
+		first = second;
+		insql(word, doff, second_dest);
+		checksum += word;
+		stq_u(partial_dest | second_dest, dst);
+		carry = checksum < word;
+		insqh(word, doff, partial_dest);
+		dst++;
+	}
+	len += doff;
+	checksum += carry;
+	if (len >= 0) {
+		unsigned long second, word;
+		unsigned long second_dest;
+
+		err |= __get_user_u(second, lastsrc);
+		extql(first, soff, word);
+		extqh(second, soff, first);
+		word |= first;
+		first = second;
+		mskql(word, len-doff, word);
+		checksum += word;
+		insql(word, doff, second_dest);
+		carry = checksum < word;
+		stq_u(partial_dest | second_dest, dst);
+		if (len) {
+			ldq_u(second_dest, dst+1);
+			insqh(word, doff, partial_dest);
+			mskqh(second_dest, len, second_dest);
+			stq_u(partial_dest | second_dest, dst+1);
+		}
+		checksum += carry;
+	} else {
+		unsigned long second, word;
+		unsigned long second_dest;
+
+		err |= __get_user_u(second, lastsrc);
+		extql(first, soff, word);
+		extqh(second, soff, first);
+		word |= first;
+		ldq_u(second_dest, dst);
+		mskql(word, len-doff, word);
+		checksum += word;
+		mskqh(second_dest, len, second_dest);
+		carry = checksum < word;
+		insql(word, doff, word);
+		stq_u(partial_dest | word | second_dest, dst);
+		checksum += carry;
+	}
+	if (err && errp) *errp = err;
+	return checksum;
+}
+
+__wsum
+csum_partial_copy_from_user(const void __user *src, void *dst, int len,
+			       __wsum sum, int *errp)
+{
+	unsigned long checksum = (__force u32) sum;
+	unsigned long soff = 7 & (unsigned long) src;
+	unsigned long doff = 7 & (unsigned long) dst;
+
+	if (len) {
+		if (!access_ok(VERIFY_READ, src, len)) {
+			if (errp) *errp = -EFAULT;
+			memset(dst, 0, len);
+			return sum;
+		}
+		if (!doff) {
+			if (!soff)
+				checksum = csum_partial_cfu_aligned(
+					(const unsigned long __user *) src,
+					(unsigned long *) dst,
+					len-8, checksum, errp);
+			else
+				checksum = csum_partial_cfu_dest_aligned(
+					(const unsigned long __user *) src,
+					(unsigned long *) dst,
+					soff, len-8, checksum, errp);
+		} else {
+			unsigned long partial_dest;
+			ldq_u(partial_dest, dst);
+			if (!soff)
+				checksum = csum_partial_cfu_src_aligned(
+					(const unsigned long __user *) src,
+					(unsigned long *) dst,
+					doff, len-8, checksum,
+					partial_dest, errp);
+			else
+				checksum = csum_partial_cfu_unaligned(
+					(const unsigned long __user *) src,
+					(unsigned long *) dst,
+					soff, doff, len-8, checksum,
+					partial_dest, errp);
+		}
+		checksum = from64to16 (checksum);
+	}
+	return (__force __wsum)checksum;
+}
+
+__wsum
+csum_partial_copy_nocheck(const void *src, void *dst, int len, __wsum sum)
+{
+	__wsum checksum;
+	mm_segment_t oldfs = get_fs();
+	set_fs(KERNEL_DS);
+	checksum = csum_partial_copy_from_user((__force const void __user *)src,
+						dst, len, sum, NULL);
+	set_fs(oldfs);
+	return checksum;
+}

arch/alpha/lib/dbg_current.S

@@ -0,0 +1,29 @@
+/*
+ * arch/alpha/lib/dbg_current.S
+ * Contributed by Richard Henderson (rth@cygnus.com)
+ *
+ * Trap if we find current not correct.
+ */
+
+#include <asm/pal.h>
+
+	.text
+	.set noat
+
+	.globl _mcount
+	.ent _mcount
+_mcount:
+	.frame $30, 0, $28, 0
+	.prologue 0
+
+	lda	$0, -0x4000($30)
+	cmpult	$8, $30, $1
+	cmpule	$0, $30, $2
+	and	$1, $2, $3
+	bne	$3, 1f
+
+	call_pal PAL_bugchk
+
+1:	ret	$31, ($28), 1
+
+	.end _mcount

arch/alpha/lib/dbg_stackcheck.S

@@ -0,0 +1,27 @@
+/*
+ * arch/alpha/lib/dbg_stackcheck.S
+ * Contributed by Richard Henderson (rth@tamu.edu)
+ *
+ * Verify that we have not overflowed the stack.  Oops if we have.
+ */
+
+#include <asm/asm-offsets.h>
+
+	.text
+	.set noat
+
+	.align 3
+	.globl _mcount
+	.ent _mcount
+_mcount:
+	.frame $30, 0, $28, 0
+	.prologue 0
+
+	lda	$0, TASK_SIZE($8)
+	cmpult	$30, $0, $0
+	bne	$0, 1f
+	ret	($28)
+1:	stq	$31, -8($31)	# oops me, damn it.
+	br	1b
+
+	.end _mcount

arch/alpha/lib/dbg_stackkill.S

@@ -0,0 +1,35 @@
+/*
+ * arch/alpha/lib/dbg_stackkill.S
+ * Contributed by Richard Henderson (rth@cygnus.com)
+ *
+ * Clobber the balance of the kernel stack, hoping to catch
+ * uninitialized local variables in the act.
+ */
+
+#include <asm/asm-offsets.h>
+
+	.text
+	.set noat
+
+	.align 5
+	.globl _mcount
+	.ent _mcount
+_mcount:
+	.frame $30, 0, $28, 0
+	.prologue 0
+
+	ldi	$0, 0xdeadbeef
+	lda	$2, -STACK_SIZE
+	sll	$0, 32, $1
+	and	$30, $2, $2
+	or	$0, $1, $0
+	lda	$2, TASK_SIZE($2)
+	cmpult	$2, $30, $1
+	beq	$1, 2f
+1:	stq	$0, 0($2)
+	addq	$2, 8, $2
+	cmpult	$2, $30, $1
+	bne	$1, 1b
+2:	ret	($28)
+
+	.end _mcount

arch/alpha/lib/dec_and_lock.c

@@ -0,0 +1,41 @@
+/*
+ * arch/alpha/lib/dec_and_lock.c
+ *
+ * ll/sc version of atomic_dec_and_lock()
+ * 
+ */
+
+#include <linux/spinlock.h>
+#include <linux/atomic.h>
+
+  asm (".text					\n\
+	.global _atomic_dec_and_lock		\n\
+	.ent _atomic_dec_and_lock		\n\
+	.align	4				\n\
+_atomic_dec_and_lock:				\n\
+	.prologue 0				\n\
+1:	ldl_l	$1, 0($16)			\n\
+	subl	$1, 1, $1			\n\
+	beq	$1, 2f				\n\
+	stl_c	$1, 0($16)			\n\
+	beq	$1, 4f				\n\
+	mb					\n\
+	clr	$0				\n\
+	ret					\n\
+2:	br	$29, 3f				\n\
+3:	ldgp	$29, 0($29)			\n\
+	br	$atomic_dec_and_lock_1..ng	\n\
+	.subsection 2				\n\
+4:	br	1b				\n\
+	.previous				\n\
+	.end _atomic_dec_and_lock");
+
+static int __used atomic_dec_and_lock_1(atomic_t *atomic, spinlock_t *lock)
+{
+	/* Slow path */
+	spin_lock(lock);
+	if (atomic_dec_and_test(atomic))
+		return 1;
+	spin_unlock(lock);
+	return 0;
+}

arch/alpha/lib/divide.S

@@ -0,0 +1,195 @@
+/*
+ * arch/alpha/lib/divide.S
+ *
+ * (C) 1995 Linus Torvalds
+ *
+ * Alpha division..
+ */
+
+/*
+ * The alpha chip doesn't provide hardware division, so we have to do it
+ * by hand.  The compiler expects the functions
+ *
+ *	__divqu: 64-bit unsigned long divide
+ *	__remqu: 64-bit unsigned long remainder
+ *	__divqs/__remqs: signed 64-bit
+ *	__divlu/__remlu: unsigned 32-bit
+ *	__divls/__remls: signed 32-bit
+ *
+ * These are not normal C functions: instead of the normal
+ * calling sequence, these expect their arguments in registers
+ * $24 and $25, and return the result in $27. Register $28 may
+ * be clobbered (assembly temporary), anything else must be saved. 
+ *
+ * In short: painful.
+ *
+ * This is a rather simple bit-at-a-time algorithm: it's very good
+ * at dividing random 64-bit numbers, but the more usual case where
+ * the divisor is small is handled better by the DEC algorithm
+ * using lookup tables. This uses much less memory, though, and is
+ * nicer on the cache.. Besides, I don't know the copyright status
+ * of the DEC code.
+ */
+
+/*
+ * My temporaries:
+ *	$0 - current bit
+ *	$1 - shifted divisor
+ *	$2 - modulus/quotient
+ *
+ *	$23 - return address
+ *	$24 - dividend
+ *	$25 - divisor
+ *
+ *	$27 - quotient/modulus
+ *	$28 - compare status
+ */
+
+#define halt .long 0
+
+/*
+ * Select function type and registers
+ */
+#define mask	$0
+#define divisor	$1
+#define compare $28
+#define tmp1	$3
+#define tmp2	$4
+
+#ifdef DIV
+#define DIV_ONLY(x,y...) x,##y
+#define MOD_ONLY(x,y...)
+#define func(x) __div##x
+#define modulus $2
+#define quotient $27
+#define GETSIGN(x) xor $24,$25,x
+#define STACK 48
+#else
+#define DIV_ONLY(x,y...)
+#define MOD_ONLY(x,y...) x,##y
+#define func(x) __rem##x
+#define modulus $27
+#define quotient $2
+#define GETSIGN(x) bis $24,$24,x
+#define STACK 32
+#endif
+
+/*
+ * For 32-bit operations, we need to extend to 64-bit
+ */
+#ifdef INTSIZE
+#define ufunction func(lu)
+#define sfunction func(l)
+#define LONGIFY(x) zapnot x,15,x
+#define SLONGIFY(x) addl x,0,x
+#else
+#define ufunction func(qu)
+#define sfunction func(q)
+#define LONGIFY(x)
+#define SLONGIFY(x)
+#endif
+
+.set noat
+.align	3
+.globl	ufunction
+.ent	ufunction
+ufunction:
+	subq	$30,STACK,$30
+	.frame	$30,STACK,$23
+	.prologue 0
+
+7:	stq	$1, 0($30)
+	bis	$25,$25,divisor
+	stq	$2, 8($30)
+	bis	$24,$24,modulus
+	stq	$0,16($30)
+	bis	$31,$31,quotient
+	LONGIFY(divisor)
+	stq	tmp1,24($30)
+	LONGIFY(modulus)
+	bis	$31,1,mask
+	DIV_ONLY(stq tmp2,32($30))
+	beq	divisor, 9f			/* div by zero */
+
+#ifdef INTSIZE
+	/*
+	 * shift divisor left, using 3-bit shifts for
+	 * 32-bit divides as we can't overflow. Three-bit
+	 * shifts will result in looping three times less
+	 * here, but can result in two loops more later.
+	 * Thus using a large shift isn't worth it (and
+	 * s8add pairs better than a sll..)
+	 */
+1:	cmpult	divisor,modulus,compare
+	s8addq	divisor,$31,divisor
+	s8addq	mask,$31,mask
+	bne	compare,1b
+#else
+1:	cmpult	divisor,modulus,compare
+	blt     divisor, 2f
+	addq	divisor,divisor,divisor
+	addq	mask,mask,mask
+	bne	compare,1b
+	unop
+#endif
+
+	/* ok, start to go right again.. */
+2:	DIV_ONLY(addq quotient,mask,tmp2)
+	srl	mask,1,mask
+	cmpule	divisor,modulus,compare
+	subq	modulus,divisor,tmp1
+	DIV_ONLY(cmovne compare,tmp2,quotient)
+	srl	divisor,1,divisor
+	cmovne	compare,tmp1,modulus
+	bne	mask,2b
+
+9:	ldq	$1, 0($30)
+	ldq	$2, 8($30)
+	ldq	$0,16($30)
+	ldq	tmp1,24($30)
+	DIV_ONLY(ldq tmp2,32($30))
+	addq	$30,STACK,$30
+	ret	$31,($23),1
+	.end	ufunction
+
+/*
+ * Uhh.. Ugly signed division. I'd rather not have it at all, but
+ * it's needed in some circumstances. There are different ways to
+ * handle this, really. This does:
+ * 	-a / b = a / -b = -(a / b)
+ *	-a % b = -(a % b)
+ *	a % -b = a % b
+ * which is probably not the best solution, but at least should
+ * have the property that (x/y)*y + (x%y) = x.
+ */
+.align 3
+.globl	sfunction
+.ent	sfunction
+sfunction:
+	subq	$30,STACK,$30
+	.frame	$30,STACK,$23
+	.prologue 0
+	bis	$24,$25,$28
+	SLONGIFY($28)
+	bge	$28,7b
+	stq	$24,0($30)
+	subq	$31,$24,$28
+	stq	$25,8($30)
+	cmovlt	$24,$28,$24	/* abs($24) */
+	stq	$23,16($30)
+	subq	$31,$25,$28
+	stq	tmp1,24($30)
+	cmovlt	$25,$28,$25	/* abs($25) */
+	unop
+	bsr	$23,ufunction
+	ldq	$24,0($30)
+	ldq	$25,8($30)
+	GETSIGN($28)
+	subq	$31,$27,tmp1
+	SLONGIFY($28)
+	ldq	$23,16($30)
+	cmovlt	$28,tmp1,$27
+	ldq	tmp1,24($30)
+	addq	$30,STACK,$30
+	ret	$31,($23),1
+	.end	sfunction

arch/alpha/lib/ev6-clear_page.S

@@ -0,0 +1,54 @@
+/*
+ * arch/alpha/lib/ev6-clear_page.S
+ *
+ * Zero an entire page.
+ */
+
+        .text
+        .align 4
+        .global clear_page
+        .ent clear_page
+clear_page:
+        .prologue 0
+
+	lda	$0,128
+	lda	$1,125
+	addq	$16,64,$2
+	addq	$16,128,$3
+
+	addq	$16,192,$17
+	wh64	($16)
+	wh64	($2)
+	wh64	($3)
+
+1:	wh64	($17)
+	stq	$31,0($16)
+	subq	$0,1,$0
+	subq	$1,1,$1
+
+	stq	$31,8($16)
+	stq	$31,16($16)
+	addq	$17,64,$2
+	nop
+
+	stq	$31,24($16)
+	stq	$31,32($16)
+	cmovgt	$1,$2,$17
+	nop
+
+	stq	$31,40($16)
+	stq	$31,48($16)
+	nop
+	nop
+
+	stq	$31,56($16)
+	addq	$16,64,$16
+	nop
+	bne	$0,1b
+
+	ret
+	nop
+	nop
+	nop
+
+	.end clear_page

arch/alpha/lib/ev6-clear_user.S

@@ -0,0 +1,225 @@
+/*
+ * arch/alpha/lib/ev6-clear_user.S
+ * 21264 version contributed by Rick Gorton <rick.gorton@alpha-processor.com>
+ *
+ * Zero user space, handling exceptions as we go.
+ *
+ * We have to make sure that $0 is always up-to-date and contains the
+ * right "bytes left to zero" value (and that it is updated only _after_
+ * a successful copy).  There is also some rather minor exception setup
+ * stuff.
+ *
+ * NOTE! This is not directly C-callable, because the calling semantics
+ * are different:
+ *
+ * Inputs:
+ *	length in $0
+ *	destination address in $6
+ *	exception pointer in $7
+ *	return address in $28 (exceptions expect it there)
+ *
+ * Outputs:
+ *	bytes left to copy in $0
+ *
+ * Clobbers:
+ *	$1,$2,$3,$4,$5,$6
+ *
+ * Much of the information about 21264 scheduling/coding comes from:
+ *	Compiler Writer's Guide for the Alpha 21264
+ *	abbreviated as 'CWG' in other comments here
+ *	ftp.digital.com/pub/Digital/info/semiconductor/literature/dsc-library.html
+ * Scheduling notation:
+ *	E	- either cluster
+ *	U	- upper subcluster; U0 - subcluster U0; U1 - subcluster U1
+ *	L	- lower subcluster; L0 - subcluster L0; L1 - subcluster L1
+ * Try not to change the actual algorithm if possible for consistency.
+ * Determining actual stalls (other than slotting) doesn't appear to be easy to do.
+ * From perusing the source code context where this routine is called, it is
+ * a fair assumption that significant fractions of entire pages are zeroed, so
+ * it's going to be worth the effort to hand-unroll a big loop, and use wh64.
+ * ASSUMPTION:
+ *	The believed purpose of only updating $0 after a store is that a signal
+ *	may come along during the execution of this chunk of code, and we don't
+ *	want to leave a hole (and we also want to avoid repeating lots of work)
+ */
+
+/* Allow an exception for an insn; exit if we get one.  */
+#define EX(x,y...)			\
+	99: x,##y;			\
+	.section __ex_table,"a";	\
+	.long 99b - .;			\
+	lda $31, $exception-99b($31); 	\
+	.previous
+
+	.set noat
+	.set noreorder
+	.align 4
+
+	.globl __do_clear_user
+	.ent __do_clear_user
+	.frame	$30, 0, $28
+	.prologue 0
+
+				# Pipeline info : Slotting & Comments
+__do_clear_user:
+	and	$6, 7, $4	# .. E  .. ..	: find dest head misalignment
+	beq	$0, $zerolength # U  .. .. ..	:  U L U L
+
+	addq	$0, $4, $1	# .. .. .. E	: bias counter
+	and	$1, 7, $2	# .. .. E  ..	: number of misaligned bytes in tail
+# Note - we never actually use $2, so this is a moot computation
+# and we can rewrite this later...
+	srl	$1, 3, $1	# .. E  .. ..	: number of quadwords to clear
+	beq	$4, $headalign	# U  .. .. ..	: U L U L
+
+/*
+ * Head is not aligned.  Write (8 - $4) bytes to head of destination
+ * This means $6 is known to be misaligned
+ */
+	EX( ldq_u $5, 0($6) )	# .. .. .. L	: load dst word to mask back in
+	beq	$1, $onebyte	# .. .. U  ..	: sub-word store?
+	mskql	$5, $6, $5	# .. U  .. ..	: take care of misaligned head
+	addq	$6, 8, $6	# E  .. .. .. 	: L U U L
+
+	EX( stq_u $5, -8($6) )	# .. .. .. L	:
+	subq	$1, 1, $1	# .. .. E  ..	:
+	addq	$0, $4, $0	# .. E  .. ..	: bytes left -= 8 - misalignment
+	subq	$0, 8, $0	# E  .. .. ..	: U L U L
+
+	.align	4
+/*
+ * (The .align directive ought to be a moot point)
+ * values upon initial entry to the loop
+ * $1 is number of quadwords to clear (zero is a valid value)
+ * $2 is number of trailing bytes (0..7) ($2 never used...)
+ * $6 is known to be aligned 0mod8
+ */
+$headalign:
+	subq	$1, 16, $4	# .. .. .. E	: If < 16, we can not use the huge loop
+	and	$6, 0x3f, $2	# .. .. E  ..	: Forward work for huge loop
+	subq	$2, 0x40, $3	# .. E  .. ..	: bias counter (huge loop)
+	blt	$4, $trailquad	# U  .. .. ..	: U L U L
+
+/*
+ * We know that we're going to do at least 16 quads, which means we are
+ * going to be able to use the large block clear loop at least once.
+ * Figure out how many quads we need to clear before we are 0mod64 aligned
+ * so we can use the wh64 instruction.
+ */
+
+	nop			# .. .. .. E
+	nop			# .. .. E  ..
+	nop			# .. E  .. ..
+	beq	$3, $bigalign	# U  .. .. ..	: U L U L : Aligned 0mod64
+
+$alignmod64:
+	EX( stq_u $31, 0($6) )	# .. .. .. L
+	addq	$3, 8, $3	# .. .. E  ..
+	subq	$0, 8, $0	# .. E  .. ..
+	nop			# E  .. .. ..	: U L U L
+
+	nop			# .. .. .. E
+	subq	$1, 1, $1	# .. .. E  ..
+	addq	$6, 8, $6	# .. E  .. ..
+	blt	$3, $alignmod64	# U  .. .. ..	: U L U L
+
+$bigalign:
+/*
+ * $0 is the number of bytes left
+ * $1 is the number of quads left
+ * $6 is aligned 0mod64
+ * we know that we'll be taking a minimum of one trip through
+ * CWG Section 3.7.6: do not expect a sustained store rate of > 1/cycle
+ * We are _not_ going to update $0 after every single store.  That
+ * would be silly, because there will be cross-cluster dependencies
+ * no matter how the code is scheduled.  By doing it in slightly
+ * staggered fashion, we can still do this loop in 5 fetches
+ * The worse case will be doing two extra quads in some future execution,
+ * in the event of an interrupted clear.
+ * Assumes the wh64 needs to be for 2 trips through the loop in the future
+ * The wh64 is issued on for the starting destination address for trip +2
+ * through the loop, and if there are less than two trips left, the target
+ * address will be for the current trip.
+ */
+	nop			# E :
+	nop			# E :
+	nop			# E :
+	bis	$6,$6,$3	# E : U L U L : Initial wh64 address is dest
+	/* This might actually help for the current trip... */
+
+$do_wh64:
+	wh64	($3)		# .. .. .. L1	: memory subsystem hint
+	subq	$1, 16, $4	# .. .. E  ..	: Forward calculation - repeat the loop?
+	EX( stq_u $31, 0($6) )	# .. L  .. ..
+	subq	$0, 8, $0	# E  .. .. ..	: U L U L
+
+	addq	$6, 128, $3	# E : Target address of wh64
+	EX( stq_u $31, 8($6) )	# L :
+	EX( stq_u $31, 16($6) )	# L :
+	subq	$0, 16, $0	# E : U L L U
+
+	nop			# E :
+	EX( stq_u $31, 24($6) )	# L :
+	EX( stq_u $31, 32($6) )	# L :
+	subq	$0, 168, $5	# E : U L L U : two trips through the loop left?
+	/* 168 = 192 - 24, since we've already completed some stores */
+
+	subq	$0, 16, $0	# E :
+	EX( stq_u $31, 40($6) )	# L :
+	EX( stq_u $31, 48($6) )	# L :
+	cmovlt	$5, $6, $3	# E : U L L U : Latency 2, extra mapping cycle
+
+	subq	$1, 8, $1	# E :
+	subq	$0, 16, $0	# E :
+	EX( stq_u $31, 56($6) )	# L :
+	nop			# E : U L U L
+
+	nop			# E :
+	subq	$0, 8, $0	# E :
+	addq	$6, 64, $6	# E :
+	bge	$4, $do_wh64	# U : U L U L
+
+$trailquad:
+	# zero to 16 quadwords left to store, plus any trailing bytes
+	# $1 is the number of quadwords left to go.
+	# 
+	nop			# .. .. .. E
+	nop			# .. .. E  ..
+	nop			# .. E  .. ..
+	beq	$1, $trailbytes	# U  .. .. ..	: U L U L : Only 0..7 bytes to go
+
+$onequad:
+	EX( stq_u $31, 0($6) )	# .. .. .. L
+	subq	$1, 1, $1	# .. .. E  ..
+	subq	$0, 8, $0	# .. E  .. ..
+	nop			# E  .. .. ..	: U L U L
+
+	nop			# .. .. .. E
+	nop			# .. .. E  ..
+	addq	$6, 8, $6	# .. E  .. ..
+	bgt	$1, $onequad	# U  .. .. ..	: U L U L
+
+	# We have an unknown number of bytes left to go.
+$trailbytes:
+	nop			# .. .. .. E
+	nop			# .. .. E  ..
+	nop			# .. E  .. ..
+	beq	$0, $zerolength	# U  .. .. ..	: U L U L
+
+	# $0 contains the number of bytes left to copy (0..31)
+	# so we will use $0 as the loop counter
+	# We know for a fact that $0 > 0 zero due to previous context
+$onebyte:
+	EX( stb $31, 0($6) )	# .. .. .. L
+	subq	$0, 1, $0	# .. .. E  ..	:
+	addq	$6, 1, $6	# .. E  .. ..	:
+	bgt	$0, $onebyte	# U  .. .. ..	: U L U L
+
+$zerolength:
+$exception:			# Destination for exception recovery(?)
+	nop			# .. .. .. E	:
+	nop			# .. .. E  ..	:
+	nop			# .. E  .. ..	:
+	ret	$31, ($28), 1	# L0 .. .. ..	: L U L U
+	.end __do_clear_user
+

arch/alpha/lib/ev6-copy_page.S

@@ -0,0 +1,203 @@
+/*
+ * arch/alpha/lib/ev6-copy_page.S
+ *
+ * Copy an entire page.
+ */
+
+/* The following comparison of this routine vs the normal copy_page.S
+   was written by an unnamed ev6 hardware designer and forwarded to me
+   via Steven Hobbs <hobbs@steven.zko.dec.com>.
+ 
+   First Problem: STQ overflows.
+   -----------------------------
+
+	It would be nice if EV6 handled every resource overflow efficiently,
+	but for some it doesn't.  Including store queue overflows.  It causes
+	a trap and a restart of the pipe.
+
+	To get around this we sometimes use (to borrow a term from a VSSAD
+	researcher) "aeration".  The idea is to slow the rate at which the
+	processor receives valid instructions by inserting nops in the fetch
+	path.  In doing so, you can prevent the overflow and actually make
+	the code run faster.  You can, of course, take advantage of the fact
+	that the processor can fetch at most 4 aligned instructions per cycle.
+
+	I inserted enough nops to force it to take 10 cycles to fetch the
+	loop code.  In theory, EV6 should be able to execute this loop in
+	9 cycles but I was not able to get it to run that fast -- the initial
+	conditions were such that I could not reach this optimum rate on
+	(chaotic) EV6.  I wrote the code such that everything would issue
+	in order. 
+
+   Second Problem: Dcache index matches.
+   -------------------------------------
+
+	If you are going to use this routine on random aligned pages, there
+	is a 25% chance that the pages will be at the same dcache indices.
+	This results in many nasty memory traps without care.
+
+	The solution is to schedule the prefetches to avoid the memory
+	conflicts.  I schedule the wh64 prefetches farther ahead of the
+	read prefetches to avoid this problem.
+
+   Third Problem: Needs more prefetching.
+   --------------------------------------
+
+	In order to improve the code I added deeper prefetching to take the
+	most advantage of EV6's bandwidth.
+
+	I also prefetched the read stream. Note that adding the read prefetch
+	forced me to add another cycle to the inner-most kernel - up to 11
+	from the original 8 cycles per iteration.  We could improve performance
+	further by unrolling the loop and doing multiple prefetches per cycle.
+
+   I think that the code below will be very robust and fast code for the
+   purposes of copying aligned pages.  It is slower when both source and
+   destination pages are in the dcache, but it is my guess that this is
+   less important than the dcache miss case.  */
+
+
+	.text
+	.align 4
+	.global copy_page
+	.ent copy_page
+copy_page:
+	.prologue 0
+
+	/* Prefetch 5 read cachelines; write-hint 10 cache lines.  */
+	wh64	($16)
+	ldl	$31,0($17)
+	ldl	$31,64($17)
+	lda	$1,1*64($16)
+
+	wh64	($1)
+	ldl	$31,128($17)
+	ldl	$31,192($17)
+	lda	$1,2*64($16)
+
+	wh64	($1)
+	ldl	$31,256($17)
+	lda	$18,118
+	lda	$1,3*64($16)
+
+	wh64	($1)
+	nop
+	lda	$1,4*64($16)
+	lda	$2,5*64($16)
+
+	wh64	($1)
+	wh64	($2)
+	lda	$1,6*64($16)
+	lda	$2,7*64($16)
+
+	wh64	($1)
+	wh64	($2)
+	lda	$1,8*64($16)
+	lda	$2,9*64($16)
+
+	wh64	($1)
+	wh64	($2)
+	lda	$19,10*64($16)
+	nop
+
+	/* Main prefetching/write-hinting loop.  */
+1:	ldq	$0,0($17)
+	ldq	$1,8($17)
+	unop
+	unop
+
+	unop
+	unop
+	ldq	$2,16($17)
+	ldq	$3,24($17)
+
+	ldq	$4,32($17)
+	ldq	$5,40($17)
+	unop
+	unop
+
+	unop
+	unop
+	ldq	$6,48($17)
+	ldq	$7,56($17)
+
+	ldl	$31,320($17)
+	unop
+	unop
+	unop
+
+	/* This gives the extra cycle of aeration above the minimum.  */
+	unop			
+	unop
+	unop
+	unop
+
+	wh64	($19)
+	unop
+	unop
+	unop
+
+	stq	$0,0($16)
+	subq	$18,1,$18
+	stq	$1,8($16)
+	unop
+
+	unop
+	stq	$2,16($16)
+	addq	$17,64,$17
+	stq	$3,24($16)
+
+	stq	$4,32($16)
+	stq	$5,40($16)
+	addq	$19,64,$19
+	unop
+
+	stq	$6,48($16)
+	stq	$7,56($16)
+	addq	$16,64,$16
+	bne	$18, 1b
+
+	/* Prefetch the final 5 cache lines of the read stream.  */
+	lda	$18,10
+	ldl	$31,320($17)
+	ldl	$31,384($17)
+	ldl	$31,448($17)
+
+	ldl	$31,512($17)
+	ldl	$31,576($17)
+	nop
+	nop
+
+	/* Non-prefetching, non-write-hinting cleanup loop for the
+	   final 10 cache lines.  */
+2:	ldq	$0,0($17)
+	ldq	$1,8($17)
+	ldq	$2,16($17)
+	ldq	$3,24($17)
+
+	ldq	$4,32($17)
+	ldq	$5,40($17)
+	ldq	$6,48($17)
+	ldq	$7,56($17)
+
+	stq	$0,0($16)
+	subq	$18,1,$18
+	stq	$1,8($16)
+	addq	$17,64,$17
+
+	stq	$2,16($16)
+	stq	$3,24($16)
+	stq	$4,32($16)
+	stq	$5,40($16)
+
+	stq	$6,48($16)
+	stq	$7,56($16)
+	addq	$16,64,$16
+	bne	$18, 2b
+
+	ret
+	nop
+	unop
+	nop
+
+	.end copy_page

arch/alpha/lib/ev6-copy_user.S

@@ -0,0 +1,259 @@
+/*
+ * arch/alpha/lib/ev6-copy_user.S
+ *
+ * 21264 version contributed by Rick Gorton <rick.gorton@alpha-processor.com>
+ *
+ * Copy to/from user space, handling exceptions as we go..  This
+ * isn't exactly pretty.
+ *
+ * This is essentially the same as "memcpy()", but with a few twists.
+ * Notably, we have to make sure that $0 is always up-to-date and
+ * contains the right "bytes left to copy" value (and that it is updated
+ * only _after_ a successful copy). There is also some rather minor
+ * exception setup stuff..
+ *
+ * NOTE! This is not directly C-callable, because the calling semantics are
+ * different:
+ *
+ * Inputs:
+ *	length in $0
+ *	destination address in $6
+ *	source address in $7
+ *	return address in $28
+ *
+ * Outputs:
+ *	bytes left to copy in $0
+ *
+ * Clobbers:
+ *	$1,$2,$3,$4,$5,$6,$7
+ *
+ * Much of the information about 21264 scheduling/coding comes from:
+ *	Compiler Writer's Guide for the Alpha 21264
+ *	abbreviated as 'CWG' in other comments here
+ *	ftp.digital.com/pub/Digital/info/semiconductor/literature/dsc-library.html
+ * Scheduling notation:
+ *	E	- either cluster
+ *	U	- upper subcluster; U0 - subcluster U0; U1 - subcluster U1
+ *	L	- lower subcluster; L0 - subcluster L0; L1 - subcluster L1
+ */
+
+/* Allow an exception for an insn; exit if we get one.  */
+#define EXI(x,y...)			\
+	99: x,##y;			\
+	.section __ex_table,"a";	\
+	.long 99b - .;			\
+	lda $31, $exitin-99b($31);	\
+	.previous
+
+#define EXO(x,y...)			\
+	99: x,##y;			\
+	.section __ex_table,"a";	\
+	.long 99b - .;			\
+	lda $31, $exitout-99b($31);	\
+	.previous
+
+	.set noat
+	.align 4
+	.globl __copy_user
+	.ent __copy_user
+				# Pipeline info: Slotting & Comments
+__copy_user:
+	.prologue 0
+	subq $0, 32, $1		# .. E  .. ..	: Is this going to be a small copy?
+	beq $0, $zerolength	# U  .. .. ..	: U L U L
+
+	and $6,7,$3		# .. .. .. E	: is leading dest misalignment
+	ble $1, $onebyteloop	# .. .. U  ..	: 1st branch : small amount of data
+	beq $3, $destaligned	# .. U  .. ..	: 2nd (one cycle fetcher stall)
+	subq $3, 8, $3		# E  .. .. ..	: L U U L : trip counter
+/*
+ * The fetcher stall also hides the 1 cycle cross-cluster stall for $3 (L --> U)
+ * This loop aligns the destination a byte at a time
+ * We know we have at least one trip through this loop
+ */
+$aligndest:
+	EXI( ldbu $1,0($7) )	# .. .. .. L	: Keep loads separate from stores
+	addq $6,1,$6		# .. .. E  ..	: Section 3.8 in the CWG
+	addq $3,1,$3		# .. E  .. ..	:
+	nop			# E  .. .. ..	: U L U L
+
+/*
+ * the -1 is to compensate for the inc($6) done in a previous quadpack
+ * which allows us zero dependencies within either quadpack in the loop
+ */
+	EXO( stb $1,-1($6) )	# .. .. .. L	:
+	addq $7,1,$7		# .. .. E  ..	: Section 3.8 in the CWG
+	subq $0,1,$0		# .. E  .. ..	:
+	bne $3, $aligndest	# U  .. .. ..	: U L U L
+
+/*
+ * If we fell through into here, we have a minimum of 33 - 7 bytes
+ * If we arrived via branch, we have a minimum of 32 bytes
+ */
+$destaligned:
+	and $7,7,$1		# .. .. .. E	: Check _current_ source alignment
+	bic $0,7,$4		# .. .. E  ..	: number bytes as a quadword loop
+	EXI( ldq_u $3,0($7) )	# .. L  .. ..	: Forward fetch for fallthrough code
+	beq $1,$quadaligned	# U  .. .. ..	: U L U L
+
+/*
+ * In the worst case, we've just executed an ldq_u here from 0($7)
+ * and we'll repeat it once if we take the branch
+ */
+
+/* Misaligned quadword loop - not unrolled.  Leave it that way. */
+$misquad:
+	EXI( ldq_u $2,8($7) )	# .. .. .. L	:
+	subq $4,8,$4		# .. .. E  ..	:
+	extql $3,$7,$3		# .. U  .. ..	:
+	extqh $2,$7,$1		# U  .. .. ..	: U U L L
+
+	bis $3,$1,$1		# .. .. .. E	:
+	EXO( stq $1,0($6) )	# .. .. L  ..	:
+	addq $7,8,$7		# .. E  .. ..	:
+	subq $0,8,$0		# E  .. .. ..	: U L L U
+
+	addq $6,8,$6		# .. .. .. E	:
+	bis $2,$2,$3		# .. .. E  ..	:
+	nop			# .. E  .. ..	:
+	bne $4,$misquad		# U  .. .. ..	: U L U L
+
+	nop			# .. .. .. E
+	nop			# .. .. E  ..
+	nop			# .. E  .. ..
+	beq $0,$zerolength	# U  .. .. ..	: U L U L
+
+/* We know we have at least one trip through the byte loop */
+	EXI ( ldbu $2,0($7) )	# .. .. .. L	: No loads in the same quad
+	addq $6,1,$6		# .. .. E  ..	: as the store (Section 3.8 in CWG)
+	nop			# .. E  .. ..	:
+	br $31, $dirtyentry	# L0 .. .. ..	: L U U L
+/* Do the trailing byte loop load, then hop into the store part of the loop */
+
+/*
+ * A minimum of (33 - 7) bytes to do a quad at a time.
+ * Based upon the usage context, it's worth the effort to unroll this loop
+ * $0 - number of bytes to be moved
+ * $4 - number of bytes to move as quadwords
+ * $6 is current destination address
+ * $7 is current source address
+ */
+$quadaligned:
+	subq	$4, 32, $2	# .. .. .. E	: do not unroll for small stuff
+	nop			# .. .. E  ..
+	nop			# .. E  .. ..
+	blt	$2, $onequad	# U  .. .. ..	: U L U L
+
+/*
+ * There is a significant assumption here that the source and destination
+ * addresses differ by more than 32 bytes.  In this particular case, a
+ * sparsity of registers further bounds this to be a minimum of 8 bytes.
+ * But if this isn't met, then the output result will be incorrect.
+ * Furthermore, due to a lack of available registers, we really can't
+ * unroll this to be an 8x loop (which would enable us to use the wh64
+ * instruction memory hint instruction).
+ */
+$unroll4:
+	EXI( ldq $1,0($7) )	# .. .. .. L
+	EXI( ldq $2,8($7) )	# .. .. L  ..
+	subq	$4,32,$4	# .. E  .. ..
+	nop			# E  .. .. ..	: U U L L
+
+	addq	$7,16,$7	# .. .. .. E
+	EXO( stq $1,0($6) )	# .. .. L  ..
+	EXO( stq $2,8($6) )	# .. L  .. ..
+	subq	$0,16,$0	# E  .. .. ..	: U L L U
+
+	addq	$6,16,$6	# .. .. .. E
+	EXI( ldq $1,0($7) )	# .. .. L  ..
+	EXI( ldq $2,8($7) )	# .. L  .. ..
+	subq	$4, 32, $3	# E  .. .. ..	: U U L L : is there enough for another trip?
+
+	EXO( stq $1,0($6) )	# .. .. .. L
+	EXO( stq $2,8($6) )	# .. .. L  ..
+	subq	$0,16,$0	# .. E  .. ..
+	addq	$7,16,$7	# E  .. .. ..	: U L L U
+
+	nop			# .. .. .. E
+	nop			# .. .. E  ..
+	addq	$6,16,$6	# .. E  .. ..
+	bgt	$3,$unroll4	# U  .. .. ..	: U L U L
+
+	nop
+	nop
+	nop
+	beq	$4, $noquads
+
+$onequad:
+	EXI( ldq $1,0($7) )
+	subq	$4,8,$4
+	addq	$7,8,$7
+	nop
+
+	EXO( stq $1,0($6) )
+	subq	$0,8,$0
+	addq	$6,8,$6
+	bne	$4,$onequad
+
+$noquads:
+	nop
+	nop
+	nop
+	beq $0,$zerolength
+
+/*
+ * For small copies (or the tail of a larger copy), do a very simple byte loop.
+ * There's no point in doing a lot of complex alignment calculations to try to
+ * to quadword stuff for a small amount of data.
+ *	$0 - remaining number of bytes left to copy
+ *	$6 - current dest addr
+ *	$7 - current source addr
+ */
+
+$onebyteloop:
+	EXI ( ldbu $2,0($7) )	# .. .. .. L	: No loads in the same quad
+	addq $6,1,$6		# .. .. E  ..	: as the store (Section 3.8 in CWG)
+	nop			# .. E  .. ..	:
+	nop			# E  .. .. ..	: U L U L
+
+$dirtyentry:
+/*
+ * the -1 is to compensate for the inc($6) done in a previous quadpack
+ * which allows us zero dependencies within either quadpack in the loop
+ */
+	EXO ( stb $2,-1($6) )	# .. .. .. L	:
+	addq $7,1,$7		# .. .. E  ..	: quadpack as the load
+	subq $0,1,$0		# .. E  .. ..	: change count _after_ copy
+	bgt $0,$onebyteloop	# U  .. .. ..	: U L U L
+
+$zerolength:
+$exitout:			# Destination for exception recovery(?)
+	nop			# .. .. .. E
+	nop			# .. .. E  ..
+	nop			# .. E  .. ..
+	ret $31,($28),1		# L0 .. .. ..	: L U L U
+
+$exitin:
+
+	/* A stupid byte-by-byte zeroing of the rest of the output
+	   buffer.  This cures security holes by never leaving 
+	   random kernel data around to be copied elsewhere.  */
+
+	nop
+	nop
+	nop
+	mov	$0,$1
+
+$101:
+	EXO ( stb $31,0($6) )	# L
+	subq $1,1,$1		# E
+	addq $6,1,$6		# E
+	bgt $1,$101		# U
+
+	nop
+	nop
+	nop
+	ret $31,($28),1		# L0
+
+	.end __copy_user
+

arch/alpha/lib/ev6-csum_ipv6_magic.S

@@ -0,0 +1,150 @@
+/*
+ * arch/alpha/lib/ev6-csum_ipv6_magic.S
+ * 21264 version contributed by Rick Gorton <rick.gorton@alpha-processor.com>
+ *
+ * unsigned short csum_ipv6_magic(struct in6_addr *saddr,
+ *                                struct in6_addr *daddr,
+ *                                __u32 len,
+ *                                unsigned short proto,
+ *                                unsigned int csum);
+ *
+ * Much of the information about 21264 scheduling/coding comes from:
+ *	Compiler Writer's Guide for the Alpha 21264
+ *	abbreviated as 'CWG' in other comments here
+ *	ftp.digital.com/pub/Digital/info/semiconductor/literature/dsc-library.html
+ * Scheduling notation:
+ *	E	- either cluster
+ *	U	- upper subcluster; U0 - subcluster U0; U1 - subcluster U1
+ *	L	- lower subcluster; L0 - subcluster L0; L1 - subcluster L1
+ * Try not to change the actual algorithm if possible for consistency.
+ * Determining actual stalls (other than slotting) doesn't appear to be easy to do.
+ *
+ * unsigned short csum_ipv6_magic(struct in6_addr *saddr,
+ *                                struct in6_addr *daddr,
+ *                                __u32 len,
+ *                                unsigned short proto,
+ *                                unsigned int csum);
+ *
+ * Swap <proto> (takes form 0xaabb)
+ * Then shift it left by 48, so result is:
+ *	0xbbaa0000 00000000
+ * Then turn it back into a sign extended 32-bit item
+ *	0xbbaa0000
+ *
+ * Swap <len> (an unsigned int) using Mike Burrows' 7-instruction sequence
+ * (we can't hide the 3-cycle latency of the unpkbw in the 6-instruction sequence)
+ * Assume input takes form 0xAABBCCDD
+ *
+ * Finally, original 'folding' approach is to split the long into 4 unsigned shorts
+ * add 4 ushorts, resulting in ushort/carry
+ * add carry bits + ushort --> ushort
+ * add carry bits + ushort --> ushort (in case the carry results in an overflow)
+ * Truncate to a ushort.  (took 13 instructions)
+ * From doing some testing, using the approach in checksum.c:from64to16()
+ * results in the same outcome:
+ * split into 2 uints, add those, generating a ulong
+ * add the 3 low ushorts together, generating a uint
+ * a final add of the 2 lower ushorts
+ * truncating the result.
+ *
+ * Misalignment handling added by Ivan Kokshaysky <ink@jurassic.park.msu.ru>
+ * The cost is 16 instructions (~8 cycles), including two extra loads which
+ * may cause additional delay in rare cases (load-load replay traps).
+ */
+
+	.globl csum_ipv6_magic
+	.align 4
+	.ent csum_ipv6_magic
+	.frame $30,0,$26,0
+csum_ipv6_magic:
+	.prologue 0
+
+	ldq_u	$0,0($16)	# L : Latency: 3
+	inslh	$18,7,$4	# U : 0000000000AABBCC
+	ldq_u	$1,8($16)	# L : Latency: 3
+	sll	$19,8,$7	# U : U L U L : 0x00000000 00aabb00
+
+	and	$16,7,$6	# E : src misalignment
+	ldq_u	$5,15($16)	# L : Latency: 3
+	zapnot	$20,15,$20	# U : zero extend incoming csum
+	ldq_u	$2,0($17)	# L : U L U L : Latency: 3
+
+	extql	$0,$6,$0	# U :
+	extqh	$1,$6,$22	# U :
+	ldq_u	$3,8($17)	# L : Latency: 3
+	sll	$19,24,$19	# U : U U L U : 0x000000aa bb000000
+
+	cmoveq	$6,$31,$22	# E : src aligned?
+	ldq_u	$23,15($17)	# L : Latency: 3
+	inswl	$18,3,$18	# U : 000000CCDD000000
+	addl	$19,$7,$19	# E : U L U L : <sign bits>bbaabb00
+
+	or	$0,$22,$0	# E : 1st src word complete
+	extql	$1,$6,$1	# U :
+	or	$18,$4,$18	# E : 000000CCDDAABBCC
+	extqh	$5,$6,$5	# U : L U L U
+
+	and	$17,7,$6	# E : dst misalignment
+	extql	$2,$6,$2	# U :
+	or	$1,$5,$1	# E : 2nd src word complete
+	extqh	$3,$6,$22	# U : L U L U :
+
+	cmoveq	$6,$31,$22	# E : dst aligned?
+	extql	$3,$6,$3	# U :
+	addq	$20,$0,$20	# E : begin summing the words
+	extqh	$23,$6,$23	# U : L U L U :
+
+	srl	$18,16,$4	# U : 0000000000CCDDAA
+	or	$2,$22,$2	# E : 1st dst word complete
+	zap	$19,0x3,$19	# U : <sign bits>bbaa0000
+	or	$3,$23,$3	# E : U L U L : 2nd dst word complete
+
+	cmpult	$20,$0,$0	# E :
+	addq	$20,$1,$20	# E :
+	zapnot	$18,0xa,$18	# U : 00000000DD00BB00
+	zap	$4,0xa,$4	# U : U U L L : 0000000000CC00AA
+
+	or	$18,$4,$18	# E : 00000000DDCCBBAA
+	nop			# E :
+	cmpult	$20,$1,$1	# E :
+	addq	$20,$2,$20	# E : U L U L
+
+	cmpult	$20,$2,$2	# E :
+	addq	$20,$3,$20	# E :
+	cmpult	$20,$3,$3	# E : (1 cycle stall on $20)
+	addq	$20,$18,$20	# E : U L U L (1 cycle stall on $20)
+
+	cmpult	$20,$18,$18	# E :
+	addq	$20,$19,$20	# E : (1 cycle stall on $20)
+	addq	$0,$1,$0	# E : merge the carries back into the csum
+	addq	$2,$3,$2	# E :
+
+	cmpult	$20,$19,$19	# E :
+	addq	$18,$19,$18	# E : (1 cycle stall on $19)
+	addq	$0,$2,$0	# E :
+	addq	$20,$18,$20	# E : U L U L :
+		/* (1 cycle stall on $18, 2 cycles on $20) */
+
+	addq	$0,$20,$0	# E :
+	zapnot	$0,15,$1	# U : Start folding output (1 cycle stall on $0)
+	nop			# E :
+	srl	$0,32,$0	# U : U L U L : (1 cycle stall on $0)
+
+	addq	$1,$0,$1	# E : Finished generating ulong
+	extwl	$1,2,$2		# U : ushort[1] (1 cycle stall on $1)
+	zapnot	$1,3,$0		# U : ushort[0] (1 cycle stall on $1)
+	extwl	$1,4,$1		# U : ushort[2] (1 cycle stall on $1)
+
+	addq	$0,$2,$0	# E
+	addq	$0,$1,$3	# E : Finished generating uint
+		/* (1 cycle stall on $0) */
+	extwl	$3,2,$1		# U : ushort[1] (1 cycle stall on $3)
+	nop			# E : L U L U
+
+	addq	$1,$3,$0	# E : Final carry
+	not	$0,$4		# E : complement (1 cycle stall on $0)
+	zapnot	$4,3,$0		# U : clear upper garbage bits
+		/* (1 cycle stall on $4) */
+	ret			# L0 : L U L U
+
+	.end csum_ipv6_magic

arch/alpha/lib/ev6-divide.S

@@ -0,0 +1,259 @@
+/*
+ * arch/alpha/lib/ev6-divide.S
+ *
+ * 21264 version contributed by Rick Gorton <rick.gorton@alpha-processor.com>
+ *
+ * Alpha division..
+ */
+
+/*
+ * The alpha chip doesn't provide hardware division, so we have to do it
+ * by hand.  The compiler expects the functions
+ *
+ *	__divqu: 64-bit unsigned long divide
+ *	__remqu: 64-bit unsigned long remainder
+ *	__divqs/__remqs: signed 64-bit
+ *	__divlu/__remlu: unsigned 32-bit
+ *	__divls/__remls: signed 32-bit
+ *
+ * These are not normal C functions: instead of the normal
+ * calling sequence, these expect their arguments in registers
+ * $24 and $25, and return the result in $27. Register $28 may
+ * be clobbered (assembly temporary), anything else must be saved. 
+ *
+ * In short: painful.
+ *
+ * This is a rather simple bit-at-a-time algorithm: it's very good
+ * at dividing random 64-bit numbers, but the more usual case where
+ * the divisor is small is handled better by the DEC algorithm
+ * using lookup tables. This uses much less memory, though, and is
+ * nicer on the cache.. Besides, I don't know the copyright status
+ * of the DEC code.
+ */
+
+/*
+ * My temporaries:
+ *	$0 - current bit
+ *	$1 - shifted divisor
+ *	$2 - modulus/quotient
+ *
+ *	$23 - return address
+ *	$24 - dividend
+ *	$25 - divisor
+ *
+ *	$27 - quotient/modulus
+ *	$28 - compare status
+ *
+ * Much of the information about 21264 scheduling/coding comes from:
+ *	Compiler Writer's Guide for the Alpha 21264
+ *	abbreviated as 'CWG' in other comments here
+ *	ftp.digital.com/pub/Digital/info/semiconductor/literature/dsc-library.html
+ * Scheduling notation:
+ *	E	- either cluster
+ *	U	- upper subcluster; U0 - subcluster U0; U1 - subcluster U1
+ *	L	- lower subcluster; L0 - subcluster L0; L1 - subcluster L1
+ * Try not to change the actual algorithm if possible for consistency.
+ */
+
+#define halt .long 0
+
+/*
+ * Select function type and registers
+ */
+#define mask	$0
+#define divisor	$1
+#define compare $28
+#define tmp1	$3
+#define tmp2	$4
+
+#ifdef DIV
+#define DIV_ONLY(x,y...) x,##y
+#define MOD_ONLY(x,y...)
+#define func(x) __div##x
+#define modulus $2
+#define quotient $27
+#define GETSIGN(x) xor $24,$25,x
+#define STACK 48
+#else
+#define DIV_ONLY(x,y...)
+#define MOD_ONLY(x,y...) x,##y
+#define func(x) __rem##x
+#define modulus $27
+#define quotient $2
+#define GETSIGN(x) bis $24,$24,x
+#define STACK 32
+#endif
+
+/*
+ * For 32-bit operations, we need to extend to 64-bit
+ */
+#ifdef INTSIZE
+#define ufunction func(lu)
+#define sfunction func(l)
+#define LONGIFY(x) zapnot x,15,x
+#define SLONGIFY(x) addl x,0,x
+#else
+#define ufunction func(qu)
+#define sfunction func(q)
+#define LONGIFY(x)
+#define SLONGIFY(x)
+#endif
+
+.set noat
+.align	4
+.globl	ufunction
+.ent	ufunction
+ufunction:
+	subq	$30,STACK,$30		# E :
+	.frame	$30,STACK,$23
+	.prologue 0
+
+7:	stq	$1, 0($30)		# L :
+	bis	$25,$25,divisor		# E :
+	stq	$2, 8($30)		# L : L U L U
+
+	bis	$24,$24,modulus		# E :
+	stq	$0,16($30)		# L :
+	bis	$31,$31,quotient	# E :
+	LONGIFY(divisor)		# E : U L L U
+
+	stq	tmp1,24($30)		# L :
+	LONGIFY(modulus)		# E :
+	bis	$31,1,mask		# E :
+	DIV_ONLY(stq tmp2,32($30))	# L : L U U L
+
+	beq	divisor, 9f			/* div by zero */
+	/*
+	 * In spite of the DIV_ONLY being either a non-instruction
+	 * or an actual stq, the addition of the .align directive
+	 * below ensures that label 1 is going to be nicely aligned
+	 */
+
+	.align	4
+#ifdef INTSIZE
+	/*
+	 * shift divisor left, using 3-bit shifts for
+	 * 32-bit divides as we can't overflow. Three-bit
+	 * shifts will result in looping three times less
+	 * here, but can result in two loops more later.
+	 * Thus using a large shift isn't worth it (and
+	 * s8add pairs better than a sll..)
+	 */
+1:	cmpult	divisor,modulus,compare	# E :
+	s8addq	divisor,$31,divisor	# E :
+	s8addq	mask,$31,mask		# E :
+	bne	compare,1b		# U : U L U L
+#else
+1:	cmpult	divisor,modulus,compare	# E :
+	nop				# E :
+	nop				# E :
+	blt     divisor, 2f		# U : U L U L
+
+	addq	divisor,divisor,divisor	# E :
+	addq	mask,mask,mask		# E :
+	unop				# E :
+	bne	compare,1b		# U : U L U L
+#endif
+
+	/* ok, start to go right again.. */
+2:
+	/*
+	 * Keep things nicely bundled... use a nop instead of not
+	 * having an instruction for DIV_ONLY
+	 */
+#ifdef DIV
+	DIV_ONLY(addq quotient,mask,tmp2) # E :
+#else
+	nop				# E :
+#endif
+	srl	mask,1,mask		# U :
+	cmpule	divisor,modulus,compare	# E :
+	subq	modulus,divisor,tmp1	# E :
+
+#ifdef DIV
+	DIV_ONLY(cmovne compare,tmp2,quotient)	# E : Latency 2, extra map slot
+	nop				# E : as part of the cmovne
+	srl	divisor,1,divisor	# U :
+	nop				# E : L U L U
+
+	nop				# E :
+	cmovne	compare,tmp1,modulus	# E : Latency 2, extra map slot
+	nop				# E : as part of the cmovne
+	bne	mask,2b			# U : U L U L
+#else
+	srl	divisor,1,divisor	# U :
+	cmovne	compare,tmp1,modulus	# E : Latency 2, extra map slot
+	nop				# E : as part of the cmovne
+	bne	mask,2b			# U : U L L U
+#endif
+
+9:	ldq	$1, 0($30)		# L :
+	ldq	$2, 8($30)		# L :
+	nop				# E :
+	nop				# E : U U L L
+
+	ldq	$0,16($30)		# L :
+	ldq	tmp1,24($30)		# L :
+	nop				# E :
+	nop				# E :
+
+#ifdef DIV
+	DIV_ONLY(ldq tmp2,32($30))	# L :
+#else
+	nop				# E :
+#endif
+	addq	$30,STACK,$30		# E :
+	ret	$31,($23),1		# L0 : L U U L
+	.end	ufunction
+
+/*
+ * Uhh.. Ugly signed division. I'd rather not have it at all, but
+ * it's needed in some circumstances. There are different ways to
+ * handle this, really. This does:
+ * 	-a / b = a / -b = -(a / b)
+ *	-a % b = -(a % b)
+ *	a % -b = a % b
+ * which is probably not the best solution, but at least should
+ * have the property that (x/y)*y + (x%y) = x.
+ */
+.align 4
+.globl	sfunction
+.ent	sfunction
+sfunction:
+	subq	$30,STACK,$30		# E :
+	.frame	$30,STACK,$23
+	.prologue 0
+	bis	$24,$25,$28		# E :
+	SLONGIFY($28)			# E :
+	bge	$28,7b			# U :
+
+	stq	$24,0($30)		# L :
+	subq	$31,$24,$28		# E :
+	stq	$25,8($30)		# L :
+	nop				# E : U L U L
+
+	cmovlt	$24,$28,$24	/* abs($24) */ # E : Latency 2, extra map slot
+	nop				# E : as part of the cmov
+	stq	$23,16($30)		# L :
+	subq	$31,$25,$28		# E : U L U L
+
+	stq	tmp1,24($30)		# L :
+	cmovlt	$25,$28,$25	/* abs($25) */ # E : Latency 2, extra map slot
+	nop				# E :
+	bsr	$23,ufunction		# L0: L U L U
+
+	ldq	$24,0($30)		# L :
+	ldq	$25,8($30)		# L :
+	GETSIGN($28)			# E :
+	subq	$31,$27,tmp1		# E : U U L L
+
+	SLONGIFY($28)			# E :
+	ldq	$23,16($30)		# L :
+	cmovlt	$28,tmp1,$27		# E : Latency 2, extra map slot
+	nop				# E : U L L U : as part of the cmov
+
+	ldq	tmp1,24($30)		# L :
+	nop				# E : as part of the cmov
+	addq	$30,STACK,$30		# E :
+	ret	$31,($23),1		# L0 : L U U L
+	.end	sfunction

arch/alpha/lib/ev6-memchr.S

@@ -0,0 +1,191 @@
+/*
+ * arch/alpha/lib/ev6-memchr.S
+ *
+ * 21264 version contributed by Rick Gorton <rick.gorton@alpha-processor.com>
+ *
+ * Finds characters in a memory area.  Optimized for the Alpha:
+ *
+ *    - memory accessed as aligned quadwords only
+ *    - uses cmpbge to compare 8 bytes in parallel
+ *    - does binary search to find 0 byte in last
+ *      quadword (HAKMEM needed 12 instructions to
+ *      do this instead of the 9 instructions that
+ *      binary search needs).
+ *
+ * For correctness consider that:
+ *
+ *    - only minimum number of quadwords may be accessed
+ *    - the third argument is an unsigned long
+ *
+ * Much of the information about 21264 scheduling/coding comes from:
+ *	Compiler Writer's Guide for the Alpha 21264
+ *	abbreviated as 'CWG' in other comments here
+ *	ftp.digital.com/pub/Digital/info/semiconductor/literature/dsc-library.html
+ * Scheduling notation:
+ *	E	- either cluster
+ *	U	- upper subcluster; U0 - subcluster U0; U1 - subcluster U1
+ *	L	- lower subcluster; L0 - subcluster L0; L1 - subcluster L1
+ * Try not to change the actual algorithm if possible for consistency.
+ */
+
+        .set noreorder
+        .set noat
+
+	.align	4
+	.globl memchr
+	.ent memchr
+memchr:
+	.frame $30,0,$26,0
+	.prologue 0
+
+	# Hack -- if someone passes in (size_t)-1, hoping to just
+	# search til the end of the address space, we will overflow
+	# below when we find the address of the last byte.  Given
+	# that we will never have a 56-bit address space, cropping
+	# the length is the easiest way to avoid trouble.
+	zap	$18, 0x80, $5	# U : Bound length
+	beq	$18, $not_found	# U :
+        ldq_u   $1, 0($16)	# L : load first quadword Latency=3
+	and	$17, 0xff, $17	# E : L L U U : 00000000000000ch
+
+	insbl	$17, 1, $2	# U : 000000000000ch00
+	cmpult	$18, 9, $4	# E : small (< 1 quad) string?
+	or	$2, $17, $17	# E : 000000000000chch
+        lda     $3, -1($31)	# E : U L L U
+
+	sll	$17, 16, $2	# U : 00000000chch0000
+	addq	$16, $5, $5	# E : Max search address
+	or	$2, $17, $17	# E : 00000000chchchch
+	sll	$17, 32, $2	# U : U L L U : chchchch00000000
+
+	or	$2, $17, $17	# E : chchchchchchchch
+	extql	$1, $16, $7	# U : $7 is upper bits
+	beq	$4, $first_quad	# U :
+	ldq_u	$6, -1($5)	# L : L U U L : eight or less bytes to search Latency=3
+
+	extqh	$6, $16, $6	# U : 2 cycle stall for $6
+	mov	$16, $0		# E :
+	nop			# E :
+	or	$7, $6, $1	# E : L U L U $1 = quadword starting at $16
+
+	# Deal with the case where at most 8 bytes remain to be searched
+	# in $1.  E.g.:
+	#	$18 = 6
+	#	$1 = ????c6c5c4c3c2c1
+$last_quad:
+	negq	$18, $6		# E :
+        xor	$17, $1, $1	# E :
+	srl	$3, $6, $6	# U : $6 = mask of $18 bits set
+        cmpbge  $31, $1, $2	# E : L U L U
+
+	nop
+	nop
+	and	$2, $6, $2	# E :
+        beq     $2, $not_found	# U : U L U L
+
+$found_it:
+#ifdef CONFIG_ALPHA_EV67
+	/*
+	 * Since we are guaranteed to have set one of the bits, we don't
+	 * have to worry about coming back with a 0x40 out of cttz...
+	 */
+	cttz	$2, $3		# U0 :
+	addq	$0, $3, $0	# E : All done
+	nop			# E :
+	ret			# L0 : L U L U
+#else
+	/*
+	 * Slow and clunky.  It can probably be improved.
+	 * An exercise left for others.
+	 */
+        negq    $2, $3		# E :
+        and     $2, $3, $2	# E :
+        and     $2, 0x0f, $1	# E :
+        addq    $0, 4, $3	# E :
+
+        cmoveq  $1, $3, $0	# E : Latency 2, extra map cycle
+	nop			# E : keep with cmov
+        and     $2, 0x33, $1	# E :
+        addq    $0, 2, $3	# E : U L U L : 2 cycle stall on $0
+
+        cmoveq  $1, $3, $0	# E : Latency 2, extra map cycle
+	nop			# E : keep with cmov
+        and     $2, 0x55, $1	# E :
+        addq    $0, 1, $3	# E : U L U L : 2 cycle stall on $0
+
+        cmoveq  $1, $3, $0	# E : Latency 2, extra map cycle
+	nop
+	nop
+	ret			# L0 : L U L U
+#endif
+
+	# Deal with the case where $18 > 8 bytes remain to be
+	# searched.  $16 may not be aligned.
+	.align 4
+$first_quad:
+	andnot	$16, 0x7, $0	# E :
+        insqh   $3, $16, $2	# U : $2 = 0000ffffffffffff ($16<0:2> ff)
+        xor	$1, $17, $1	# E :
+	or	$1, $2, $1	# E : U L U L $1 = ====ffffffffffff
+
+        cmpbge  $31, $1, $2	# E :
+        bne     $2, $found_it	# U :
+	# At least one byte left to process.
+	ldq	$1, 8($0)	# L :
+	subq	$5, 1, $18	# E : U L U L
+
+	addq	$0, 8, $0	# E :
+	# Make $18 point to last quad to be accessed (the
+	# last quad may or may not be partial).
+	andnot	$18, 0x7, $18	# E :
+	cmpult	$0, $18, $2	# E :
+	beq	$2, $final	# U : U L U L
+
+	# At least two quads remain to be accessed.
+
+	subq	$18, $0, $4	# E : $4 <- nr quads to be processed
+	and	$4, 8, $4	# E : odd number of quads?
+	bne	$4, $odd_quad_count # U :
+	# At least three quads remain to be accessed
+	mov	$1, $4		# E : L U L U : move prefetched value to correct reg
+
+	.align	4
+$unrolled_loop:
+	ldq	$1, 8($0)	# L : prefetch $1
+	xor	$17, $4, $2	# E :
+	cmpbge	$31, $2, $2	# E :
+	bne	$2, $found_it	# U : U L U L
+
+	addq	$0, 8, $0	# E :
+	nop			# E :
+	nop			# E :
+	nop			# E :
+
+$odd_quad_count:
+	xor	$17, $1, $2	# E :
+	ldq	$4, 8($0)	# L : prefetch $4
+	cmpbge	$31, $2, $2	# E :
+	addq	$0, 8, $6	# E :
+
+	bne	$2, $found_it	# U :
+	cmpult	$6, $18, $6	# E :
+	addq	$0, 8, $0	# E :
+	nop			# E :
+
+	bne	$6, $unrolled_loop # U :
+	mov	$4, $1		# E : move prefetched value into $1
+	nop			# E :
+	nop			# E :
+
+$final:	subq	$5, $0, $18	# E : $18 <- number of bytes left to do
+	nop			# E :
+	nop			# E :
+	bne	$18, $last_quad	# U :
+
+$not_found:
+	mov	$31, $0		# E :
+	nop			# E :
+	nop			# E :
+	ret			# L0 :
+
+        .end memchr

arch/alpha/lib/ev6-memcpy.S

@@ -0,0 +1,248 @@
+/*
+ * arch/alpha/lib/ev6-memcpy.S
+ * 21264 version by Rick Gorton <rick.gorton@alpha-processor.com>
+ *
+ * Reasonably optimized memcpy() routine for the Alpha 21264
+ *
+ *	- memory accessed as aligned quadwords only
+ *	- uses bcmpge to compare 8 bytes in parallel
+ *
+ * Much of the information about 21264 scheduling/coding comes from:
+ *	Compiler Writer's Guide for the Alpha 21264
+ *	abbreviated as 'CWG' in other comments here
+ *	ftp.digital.com/pub/Digital/info/semiconductor/literature/dsc-library.html
+ * Scheduling notation:
+ *	E	- either cluster
+ *	U	- upper subcluster; U0 - subcluster U0; U1 - subcluster U1
+ *	L	- lower subcluster; L0 - subcluster L0; L1 - subcluster L1
+ *
+ * Temp usage notes:
+ *	$1,$2,		- scratch
+ */
+
+	.set noreorder
+	.set noat
+
+	.align	4
+	.globl memcpy
+	.ent memcpy
+memcpy:
+	.frame $30,0,$26,0
+	.prologue 0
+
+	mov	$16, $0			# E : copy dest to return
+	ble	$18, $nomoredata	# U : done with the copy?
+	xor	$16, $17, $1		# E : are source and dest alignments the same?
+	and	$1, 7, $1		# E : are they the same mod 8?
+
+	bne	$1, $misaligned		# U : Nope - gotta do this the slow way
+	/* source and dest are same mod 8 address */
+	and	$16, 7, $1		# E : Are both 0mod8?
+	beq	$1, $both_0mod8		# U : Yes
+	nop				# E :
+
+	/*
+	 * source and dest are same misalignment.  move a byte at a time
+	 * until a 0mod8 alignment for both is reached.
+	 * At least one byte more to move
+	 */
+
+$head_align:
+	ldbu	$1, 0($17)		# L : grab a byte
+	subq	$18, 1, $18		# E : count--
+	addq	$17, 1, $17		# E : src++
+	stb	$1, 0($16)		# L :
+	addq	$16, 1, $16		# E : dest++
+	and	$16, 7, $1		# E : Are we at 0mod8 yet?
+	ble	$18, $nomoredata	# U : done with the copy?
+	bne	$1, $head_align		# U :
+
+$both_0mod8:
+	cmple	$18, 127, $1		# E : Can we unroll the loop?
+	bne	$1, $no_unroll		# U :
+	and	$16, 63, $1		# E : get mod64 alignment
+	beq	$1, $do_unroll		# U : no single quads to fiddle
+
+$single_head_quad:
+	ldq	$1, 0($17)		# L : get 8 bytes
+	subq	$18, 8, $18		# E : count -= 8
+	addq	$17, 8, $17		# E : src += 8
+	nop				# E :
+
+	stq	$1, 0($16)		# L : store
+	addq	$16, 8, $16		# E : dest += 8
+	and	$16, 63, $1		# E : get mod64 alignment
+	bne	$1, $single_head_quad	# U : still not fully aligned
+
+$do_unroll:
+	addq	$16, 64, $7		# E : Initial (+1 trip) wh64 address
+	cmple	$18, 127, $1		# E : Can we go through the unrolled loop?
+	bne	$1, $tail_quads		# U : Nope
+	nop				# E : 
+
+$unroll_body:
+	wh64	($7)			# L1 : memory subsystem hint: 64 bytes at
+					# ($7) are about to be over-written
+	ldq	$6, 0($17)		# L0 : bytes 0..7
+	nop				# E :
+	nop				# E :
+
+	ldq	$4, 8($17)		# L : bytes 8..15
+	ldq	$5, 16($17)		# L : bytes 16..23
+	addq	$7, 64, $7		# E : Update next wh64 address
+	nop				# E :
+
+	ldq	$3, 24($17)		# L : bytes 24..31
+	addq	$16, 64, $1		# E : fallback value for wh64
+	nop				# E :
+	nop				# E :
+
+	addq	$17, 32, $17		# E : src += 32 bytes
+	stq	$6, 0($16)		# L : bytes 0..7
+	nop				# E :
+	nop				# E :
+
+	stq	$4, 8($16)		# L : bytes 8..15
+	stq	$5, 16($16)		# L : bytes 16..23
+	subq	$18, 192, $2		# E : At least two more trips to go?
+	nop				# E :
+
+	stq	$3, 24($16)		# L : bytes 24..31
+	addq	$16, 32, $16		# E : dest += 32 bytes
+	nop				# E :
+	nop				# E :
+
+	ldq	$6, 0($17)		# L : bytes 0..7
+	ldq	$4, 8($17)		# L : bytes 8..15
+	cmovlt	$2, $1, $7		# E : Latency 2, extra map slot - Use
+					# fallback wh64 address if < 2 more trips
+	nop				# E :
+
+	ldq	$5, 16($17)		# L : bytes 16..23
+	ldq	$3, 24($17)		# L : bytes 24..31
+	addq	$16, 32, $16		# E : dest += 32
+	subq	$18, 64, $18		# E : count -= 64
+
+	addq	$17, 32, $17		# E : src += 32
+	stq	$6, -32($16)		# L : bytes 0..7
+	stq	$4, -24($16)		# L : bytes 8..15
+	cmple	$18, 63, $1		# E : At least one more trip?
+
+	stq	$5, -16($16)		# L : bytes 16..23
+	stq	$3, -8($16)		# L : bytes 24..31
+	nop				# E :
+	beq	$1, $unroll_body
+
+$tail_quads:
+$no_unroll:
+	.align 4
+	subq	$18, 8, $18		# E : At least a quad left?
+	blt	$18, $less_than_8	# U : Nope
+	nop				# E :
+	nop				# E :
+
+$move_a_quad:
+	ldq	$1, 0($17)		# L : fetch 8
+	subq	$18, 8, $18		# E : count -= 8
+	addq	$17, 8, $17		# E : src += 8
+	nop				# E :
+
+	stq	$1, 0($16)		# L : store 8
+	addq	$16, 8, $16		# E : dest += 8
+	bge	$18, $move_a_quad	# U :
+	nop				# E :
+
+$less_than_8:
+	.align 4
+	addq	$18, 8, $18		# E : add back for trailing bytes
+	ble	$18, $nomoredata	# U : All-done
+	nop				# E :
+	nop				# E :
+
+	/* Trailing bytes */
+$tail_bytes:
+	subq	$18, 1, $18		# E : count--
+	ldbu	$1, 0($17)		# L : fetch a byte
+	addq	$17, 1, $17		# E : src++
+	nop				# E :
+
+	stb	$1, 0($16)		# L : store a byte
+	addq	$16, 1, $16		# E : dest++
+	bgt	$18, $tail_bytes	# U : more to be done?
+	nop				# E :
+
+	/* branching to exit takes 3 extra cycles, so replicate exit here */
+	ret	$31, ($26), 1		# L0 :
+	nop				# E :
+	nop				# E :
+	nop				# E :
+
+$misaligned:
+	mov	$0, $4			# E : dest temp
+	and	$0, 7, $1		# E : dest alignment mod8
+	beq	$1, $dest_0mod8		# U : life doesnt totally suck
+	nop
+
+$aligndest:
+	ble	$18, $nomoredata	# U :
+	ldbu	$1, 0($17)		# L : fetch a byte
+	subq	$18, 1, $18		# E : count--
+	addq	$17, 1, $17		# E : src++
+
+	stb	$1, 0($4)		# L : store it
+	addq	$4, 1, $4		# E : dest++
+	and	$4, 7, $1		# E : dest 0mod8 yet?
+	bne	$1, $aligndest		# U : go until we are aligned.
+
+	/* Source has unknown alignment, but dest is known to be 0mod8 */
+$dest_0mod8:
+	subq	$18, 8, $18		# E : At least a quad left?
+	blt	$18, $misalign_tail	# U : Nope
+	ldq_u	$3, 0($17)		# L : seed (rotating load) of 8 bytes
+	nop				# E :
+
+$mis_quad:
+	ldq_u	$16, 8($17)		# L : Fetch next 8
+	extql	$3, $17, $3		# U : masking
+	extqh	$16, $17, $1		# U : masking
+	bis	$3, $1, $1		# E : merged bytes to store
+
+	subq	$18, 8, $18		# E : count -= 8
+	addq	$17, 8, $17		# E : src += 8
+	stq	$1, 0($4)		# L : store 8 (aligned)
+	mov	$16, $3			# E : "rotate" source data
+
+	addq	$4, 8, $4		# E : dest += 8
+	bge	$18, $mis_quad		# U : More quads to move
+	nop
+	nop
+
+$misalign_tail:
+	addq	$18, 8, $18		# E : account for tail stuff
+	ble	$18, $nomoredata	# U :
+	nop
+	nop
+
+$misalign_byte:
+	ldbu	$1, 0($17)		# L : fetch 1
+	subq	$18, 1, $18		# E : count--
+	addq	$17, 1, $17		# E : src++
+	nop				# E :
+
+	stb	$1, 0($4)		# L : store
+	addq	$4, 1, $4		# E : dest++
+	bgt	$18, $misalign_byte	# U : more to go?
+	nop
+
+
+$nomoredata:
+	ret	$31, ($26), 1		# L0 :
+	nop				# E :
+	nop				# E :
+	nop				# E :
+
+	.end memcpy
+
+/* For backwards module compatibility.  */
+__memcpy = memcpy
+.globl __memcpy

arch/alpha/lib/ev6-memset.S

@@ -0,0 +1,599 @@
+/*
+ * arch/alpha/lib/ev6-memset.S
+ *
+ * This is an efficient (and relatively small) implementation of the C library
+ * "memset()" function for the 21264 implementation of Alpha.
+ *
+ * 21264 version  contributed by Rick Gorton <rick.gorton@alpha-processor.com>
+ *
+ * Much of the information about 21264 scheduling/coding comes from:
+ *	Compiler Writer's Guide for the Alpha 21264
+ *	abbreviated as 'CWG' in other comments here
+ *	ftp.digital.com/pub/Digital/info/semiconductor/literature/dsc-library.html
+ * Scheduling notation:
+ *	E	- either cluster
+ *	U	- upper subcluster; U0 - subcluster U0; U1 - subcluster U1
+ *	L	- lower subcluster; L0 - subcluster L0; L1 - subcluster L1
+ * The algorithm for the leading and trailing quadwords remains the same,
+ * however the loop has been unrolled to enable better memory throughput,
+ * and the code has been replicated for each of the entry points: __memset
+ * and __memsetw to permit better scheduling to eliminate the stalling
+ * encountered during the mask replication.
+ * A future enhancement might be to put in a byte store loop for really
+ * small (say < 32 bytes) memset()s.  Whether or not that change would be
+ * a win in the kernel would depend upon the contextual usage.
+ * WARNING: Maintaining this is going to be more work than the above version,
+ * as fixes will need to be made in multiple places.  The performance gain
+ * is worth it.
+ */
+
+	.set noat
+	.set noreorder
+.text
+	.globl memset
+	.globl __memset
+	.globl ___memset
+	.globl __memsetw
+	.globl __constant_c_memset
+
+	.ent ___memset
+.align 5
+___memset:
+	.frame $30,0,$26,0
+	.prologue 0
+
+	/*
+	 * Serious stalling happens.  The only way to mitigate this is to
+	 * undertake a major re-write to interleave the constant materialization
+	 * with other parts of the fall-through code.  This is important, even
+	 * though it makes maintenance tougher.
+	 * Do this later.
+	 */
+	and $17,255,$1		# E : 00000000000000ch
+	insbl $17,1,$2		# U : 000000000000ch00
+	bis $16,$16,$0		# E : return value
+	ble $18,end_b		# U : zero length requested?
+
+	addq $18,$16,$6		# E : max address to write to
+	bis	$1,$2,$17	# E : 000000000000chch
+	insbl	$1,2,$3		# U : 0000000000ch0000
+	insbl	$1,3,$4		# U : 00000000ch000000
+
+	or	$3,$4,$3	# E : 00000000chch0000
+	inswl	$17,4,$5	# U : 0000chch00000000
+	xor	$16,$6,$1	# E : will complete write be within one quadword?
+	inswl	$17,6,$2	# U : chch000000000000
+
+	or	$17,$3,$17	# E : 00000000chchchch
+	or	$2,$5,$2	# E : chchchch00000000
+	bic	$1,7,$1		# E : fit within a single quadword?
+	and	$16,7,$3	# E : Target addr misalignment
+
+	or	$17,$2,$17	# E : chchchchchchchch
+	beq	$1,within_quad_b # U :
+	nop			# E :
+	beq	$3,aligned_b	# U : target is 0mod8
+
+	/*
+	 * Target address is misaligned, and won't fit within a quadword
+	 */
+	ldq_u $4,0($16)		# L : Fetch first partial
+	bis $16,$16,$5		# E : Save the address
+	insql $17,$16,$2	# U : Insert new bytes
+	subq $3,8,$3		# E : Invert (for addressing uses)
+
+	addq $18,$3,$18		# E : $18 is new count ($3 is negative)
+	mskql $4,$16,$4		# U : clear relevant parts of the quad
+	subq $16,$3,$16		# E : $16 is new aligned destination
+	bis $2,$4,$1		# E : Final bytes
+
+	nop
+	stq_u $1,0($5)		# L : Store result
+	nop
+	nop
+
+.align 4
+aligned_b:
+	/*
+	 * We are now guaranteed to be quad aligned, with at least
+	 * one partial quad to write.
+	 */
+
+	sra $18,3,$3		# U : Number of remaining quads to write
+	and $18,7,$18		# E : Number of trailing bytes to write
+	bis $16,$16,$5		# E : Save dest address
+	beq $3,no_quad_b	# U : tail stuff only
+
+	/*
+	 * it's worth the effort to unroll this and use wh64 if possible
+	 * Lifted a bunch of code from clear_user.S
+	 * At this point, entry values are:
+	 * $16	Current destination address
+	 * $5	A copy of $16
+	 * $6	The max quadword address to write to
+	 * $18	Number trailer bytes
+	 * $3	Number quads to write
+	 */
+
+	and	$16, 0x3f, $2	# E : Forward work (only useful for unrolled loop)
+	subq	$3, 16, $4	# E : Only try to unroll if > 128 bytes
+	subq	$2, 0x40, $1	# E : bias counter (aligning stuff 0mod64)
+	blt	$4, loop_b	# U :
+
+	/*
+	 * We know we've got at least 16 quads, minimum of one trip
+	 * through unrolled loop.  Do a quad at a time to get us 0mod64
+	 * aligned.
+	 */
+
+	nop			# E :
+	nop			# E :
+	nop			# E :
+	beq	$1, $bigalign_b	# U :
+
+$alignmod64_b:
+	stq	$17, 0($5)	# L :
+	subq	$3, 1, $3	# E : For consistency later
+	addq	$1, 8, $1	# E : Increment towards zero for alignment
+	addq	$5, 8, $4	# E : Initial wh64 address (filler instruction)
+
+	nop
+	nop
+	addq	$5, 8, $5	# E : Inc address
+	blt	$1, $alignmod64_b # U :
+
+$bigalign_b:
+	/*
+	 * $3 - number quads left to go
+	 * $5 - target address (aligned 0mod64)
+	 * $17 - mask of stuff to store
+	 * Scratch registers available: $7, $2, $4, $1
+	 * we know that we'll be taking a minimum of one trip through
+ 	 * CWG Section 3.7.6: do not expect a sustained store rate of > 1/cycle
+	 * Assumes the wh64 needs to be for 2 trips through the loop in the future
+	 * The wh64 is issued on for the starting destination address for trip +2
+	 * through the loop, and if there are less than two trips left, the target
+	 * address will be for the current trip.
+	 */
+
+$do_wh64_b:
+	wh64	($4)		# L1 : memory subsystem write hint
+	subq	$3, 24, $2	# E : For determining future wh64 addresses
+	stq	$17, 0($5)	# L :
+	nop			# E :
+
+	addq	$5, 128, $4	# E : speculative target of next wh64
+	stq	$17, 8($5)	# L :
+	stq	$17, 16($5)	# L :
+	addq	$5, 64, $7	# E : Fallback address for wh64 (== next trip addr)
+
+	stq	$17, 24($5)	# L :
+	stq	$17, 32($5)	# L :
+	cmovlt	$2, $7, $4	# E : Latency 2, extra mapping cycle
+	nop
+
+	stq	$17, 40($5)	# L :
+	stq	$17, 48($5)	# L :
+	subq	$3, 16, $2	# E : Repeat the loop at least once more?
+	nop
+
+	stq	$17, 56($5)	# L :
+	addq	$5, 64, $5	# E :
+	subq	$3, 8, $3	# E :
+	bge	$2, $do_wh64_b	# U :
+
+	nop
+	nop
+	nop
+	beq	$3, no_quad_b	# U : Might have finished already
+
+.align 4
+	/*
+	 * Simple loop for trailing quadwords, or for small amounts
+	 * of data (where we can't use an unrolled loop and wh64)
+	 */
+loop_b:
+	stq $17,0($5)		# L :
+	subq $3,1,$3		# E : Decrement number quads left
+	addq $5,8,$5		# E : Inc address
+	bne $3,loop_b		# U : more?
+
+no_quad_b:
+	/*
+	 * Write 0..7 trailing bytes.
+	 */
+	nop			# E :
+	beq $18,end_b		# U : All done?
+	ldq $7,0($5)		# L :
+	mskqh $7,$6,$2		# U : Mask final quad
+
+	insqh $17,$6,$4		# U : New bits
+	bis $2,$4,$1		# E : Put it all together
+	stq $1,0($5)		# L : And back to memory
+	ret $31,($26),1		# L0 :
+
+within_quad_b:
+	ldq_u $1,0($16)		# L :
+	insql $17,$16,$2	# U : New bits
+	mskql $1,$16,$4		# U : Clear old
+	bis $2,$4,$2		# E : New result
+
+	mskql $2,$6,$4		# U :
+	mskqh $1,$6,$2		# U :
+	bis $2,$4,$1		# E :
+	stq_u $1,0($16)		# L :
+
+end_b:
+	nop
+	nop
+	nop
+	ret $31,($26),1		# L0 :
+	.end ___memset
+
+	/*
+	 * This is the original body of code, prior to replication and
+	 * rescheduling.  Leave it here, as there may be calls to this
+	 * entry point.
+	 */
+.align 4
+	.ent __constant_c_memset
+__constant_c_memset:
+	.frame $30,0,$26,0
+	.prologue 0
+
+	addq $18,$16,$6		# E : max address to write to
+	bis $16,$16,$0		# E : return value
+	xor $16,$6,$1		# E : will complete write be within one quadword?
+	ble $18,end		# U : zero length requested?
+
+	bic $1,7,$1		# E : fit within a single quadword
+	beq $1,within_one_quad	# U :
+	and $16,7,$3		# E : Target addr misalignment
+	beq $3,aligned		# U : target is 0mod8
+
+	/*
+	 * Target address is misaligned, and won't fit within a quadword
+	 */
+	ldq_u $4,0($16)		# L : Fetch first partial
+	bis $16,$16,$5		# E : Save the address
+	insql $17,$16,$2	# U : Insert new bytes
+	subq $3,8,$3		# E : Invert (for addressing uses)
+
+	addq $18,$3,$18		# E : $18 is new count ($3 is negative)
+	mskql $4,$16,$4		# U : clear relevant parts of the quad
+	subq $16,$3,$16		# E : $16 is new aligned destination
+	bis $2,$4,$1		# E : Final bytes
+
+	nop
+	stq_u $1,0($5)		# L : Store result
+	nop
+	nop
+
+.align 4
+aligned:
+	/*
+	 * We are now guaranteed to be quad aligned, with at least
+	 * one partial quad to write.
+	 */
+
+	sra $18,3,$3		# U : Number of remaining quads to write
+	and $18,7,$18		# E : Number of trailing bytes to write
+	bis $16,$16,$5		# E : Save dest address
+	beq $3,no_quad		# U : tail stuff only
+
+	/*
+	 * it's worth the effort to unroll this and use wh64 if possible
+	 * Lifted a bunch of code from clear_user.S
+	 * At this point, entry values are:
+	 * $16	Current destination address
+	 * $5	A copy of $16
+	 * $6	The max quadword address to write to
+	 * $18	Number trailer bytes
+	 * $3	Number quads to write
+	 */
+
+	and	$16, 0x3f, $2	# E : Forward work (only useful for unrolled loop)
+	subq	$3, 16, $4	# E : Only try to unroll if > 128 bytes
+	subq	$2, 0x40, $1	# E : bias counter (aligning stuff 0mod64)
+	blt	$4, loop	# U :
+
+	/*
+	 * We know we've got at least 16 quads, minimum of one trip
+	 * through unrolled loop.  Do a quad at a time to get us 0mod64
+	 * aligned.
+	 */
+
+	nop			# E :
+	nop			# E :
+	nop			# E :
+	beq	$1, $bigalign	# U :
+
+$alignmod64:
+	stq	$17, 0($5)	# L :
+	subq	$3, 1, $3	# E : For consistency later
+	addq	$1, 8, $1	# E : Increment towards zero for alignment
+	addq	$5, 8, $4	# E : Initial wh64 address (filler instruction)
+
+	nop
+	nop
+	addq	$5, 8, $5	# E : Inc address
+	blt	$1, $alignmod64	# U :
+
+$bigalign:
+	/*
+	 * $3 - number quads left to go
+	 * $5 - target address (aligned 0mod64)
+	 * $17 - mask of stuff to store
+	 * Scratch registers available: $7, $2, $4, $1
+	 * we know that we'll be taking a minimum of one trip through
+ 	 * CWG Section 3.7.6: do not expect a sustained store rate of > 1/cycle
+	 * Assumes the wh64 needs to be for 2 trips through the loop in the future
+	 * The wh64 is issued on for the starting destination address for trip +2
+	 * through the loop, and if there are less than two trips left, the target
+	 * address will be for the current trip.
+	 */
+
+$do_wh64:
+	wh64	($4)		# L1 : memory subsystem write hint
+	subq	$3, 24, $2	# E : For determining future wh64 addresses
+	stq	$17, 0($5)	# L :
+	nop			# E :
+
+	addq	$5, 128, $4	# E : speculative target of next wh64
+	stq	$17, 8($5)	# L :
+	stq	$17, 16($5)	# L :
+	addq	$5, 64, $7	# E : Fallback address for wh64 (== next trip addr)
+
+	stq	$17, 24($5)	# L :
+	stq	$17, 32($5)	# L :
+	cmovlt	$2, $7, $4	# E : Latency 2, extra mapping cycle
+	nop
+
+	stq	$17, 40($5)	# L :
+	stq	$17, 48($5)	# L :
+	subq	$3, 16, $2	# E : Repeat the loop at least once more?
+	nop
+
+	stq	$17, 56($5)	# L :
+	addq	$5, 64, $5	# E :
+	subq	$3, 8, $3	# E :
+	bge	$2, $do_wh64	# U :
+
+	nop
+	nop
+	nop
+	beq	$3, no_quad	# U : Might have finished already
+
+.align 4
+	/*
+	 * Simple loop for trailing quadwords, or for small amounts
+	 * of data (where we can't use an unrolled loop and wh64)
+	 */
+loop:
+	stq $17,0($5)		# L :
+	subq $3,1,$3		# E : Decrement number quads left
+	addq $5,8,$5		# E : Inc address
+	bne $3,loop		# U : more?
+
+no_quad:
+	/*
+	 * Write 0..7 trailing bytes.
+	 */
+	nop			# E :
+	beq $18,end		# U : All done?
+	ldq $7,0($5)		# L :
+	mskqh $7,$6,$2		# U : Mask final quad
+
+	insqh $17,$6,$4		# U : New bits
+	bis $2,$4,$1		# E : Put it all together
+	stq $1,0($5)		# L : And back to memory
+	ret $31,($26),1		# L0 :
+
+within_one_quad:
+	ldq_u $1,0($16)		# L :
+	insql $17,$16,$2	# U : New bits
+	mskql $1,$16,$4		# U : Clear old
+	bis $2,$4,$2		# E : New result
+
+	mskql $2,$6,$4		# U :
+	mskqh $1,$6,$2		# U :
+	bis $2,$4,$1		# E :
+	stq_u $1,0($16)		# L :
+
+end:
+	nop
+	nop
+	nop
+	ret $31,($26),1		# L0 :
+	.end __constant_c_memset
+
+	/*
+	 * This is a replicant of the __constant_c_memset code, rescheduled
+	 * to mask stalls.  Note that entry point names also had to change
+	 */
+	.align 5
+	.ent __memsetw
+
+__memsetw:
+	.frame $30,0,$26,0
+	.prologue 0
+
+	inswl $17,0,$5		# U : 000000000000c1c2
+	inswl $17,2,$2		# U : 00000000c1c20000
+	bis $16,$16,$0		# E : return value
+	addq	$18,$16,$6	# E : max address to write to
+
+	ble $18, end_w		# U : zero length requested?
+	inswl	$17,4,$3	# U : 0000c1c200000000
+	inswl	$17,6,$4	# U : c1c2000000000000
+	xor	$16,$6,$1	# E : will complete write be within one quadword?
+
+	or	$2,$5,$2	# E : 00000000c1c2c1c2
+	or	$3,$4,$17	# E : c1c2c1c200000000
+	bic	$1,7,$1		# E : fit within a single quadword
+	and	$16,7,$3	# E : Target addr misalignment
+
+	or	$17,$2,$17	# E : c1c2c1c2c1c2c1c2
+	beq $1,within_quad_w	# U :
+	nop
+	beq $3,aligned_w	# U : target is 0mod8
+
+	/*
+	 * Target address is misaligned, and won't fit within a quadword
+	 */
+	ldq_u $4,0($16)		# L : Fetch first partial
+	bis $16,$16,$5		# E : Save the address
+	insql $17,$16,$2	# U : Insert new bytes
+	subq $3,8,$3		# E : Invert (for addressing uses)
+
+	addq $18,$3,$18		# E : $18 is new count ($3 is negative)
+	mskql $4,$16,$4		# U : clear relevant parts of the quad
+	subq $16,$3,$16		# E : $16 is new aligned destination
+	bis $2,$4,$1		# E : Final bytes
+
+	nop
+	stq_u $1,0($5)		# L : Store result
+	nop
+	nop
+
+.align 4
+aligned_w:
+	/*
+	 * We are now guaranteed to be quad aligned, with at least
+	 * one partial quad to write.
+	 */
+
+	sra $18,3,$3		# U : Number of remaining quads to write
+	and $18,7,$18		# E : Number of trailing bytes to write
+	bis $16,$16,$5		# E : Save dest address
+	beq $3,no_quad_w	# U : tail stuff only
+
+	/*
+	 * it's worth the effort to unroll this and use wh64 if possible
+	 * Lifted a bunch of code from clear_user.S
+	 * At this point, entry values are:
+	 * $16	Current destination address
+	 * $5	A copy of $16
+	 * $6	The max quadword address to write to
+	 * $18	Number trailer bytes
+	 * $3	Number quads to write
+	 */
+
+	and	$16, 0x3f, $2	# E : Forward work (only useful for unrolled loop)
+	subq	$3, 16, $4	# E : Only try to unroll if > 128 bytes
+	subq	$2, 0x40, $1	# E : bias counter (aligning stuff 0mod64)
+	blt	$4, loop_w	# U :
+
+	/*
+	 * We know we've got at least 16 quads, minimum of one trip
+	 * through unrolled loop.  Do a quad at a time to get us 0mod64
+	 * aligned.
+	 */
+
+	nop			# E :
+	nop			# E :
+	nop			# E :
+	beq	$1, $bigalign_w	# U :
+
+$alignmod64_w:
+	stq	$17, 0($5)	# L :
+	subq	$3, 1, $3	# E : For consistency later
+	addq	$1, 8, $1	# E : Increment towards zero for alignment
+	addq	$5, 8, $4	# E : Initial wh64 address (filler instruction)
+
+	nop
+	nop
+	addq	$5, 8, $5	# E : Inc address
+	blt	$1, $alignmod64_w	# U :
+
+$bigalign_w:
+	/*
+	 * $3 - number quads left to go
+	 * $5 - target address (aligned 0mod64)
+	 * $17 - mask of stuff to store
+	 * Scratch registers available: $7, $2, $4, $1
+	 * we know that we'll be taking a minimum of one trip through
+ 	 * CWG Section 3.7.6: do not expect a sustained store rate of > 1/cycle
+	 * Assumes the wh64 needs to be for 2 trips through the loop in the future
+	 * The wh64 is issued on for the starting destination address for trip +2
+	 * through the loop, and if there are less than two trips left, the target
+	 * address will be for the current trip.
+	 */
+
+$do_wh64_w:
+	wh64	($4)		# L1 : memory subsystem write hint
+	subq	$3, 24, $2	# E : For determining future wh64 addresses
+	stq	$17, 0($5)	# L :
+	nop			# E :
+
+	addq	$5, 128, $4	# E : speculative target of next wh64
+	stq	$17, 8($5)	# L :
+	stq	$17, 16($5)	# L :
+	addq	$5, 64, $7	# E : Fallback address for wh64 (== next trip addr)
+
+	stq	$17, 24($5)	# L :
+	stq	$17, 32($5)	# L :
+	cmovlt	$2, $7, $4	# E : Latency 2, extra mapping cycle
+	nop
+
+	stq	$17, 40($5)	# L :
+	stq	$17, 48($5)	# L :
+	subq	$3, 16, $2	# E : Repeat the loop at least once more?
+	nop
+
+	stq	$17, 56($5)	# L :
+	addq	$5, 64, $5	# E :
+	subq	$3, 8, $3	# E :
+	bge	$2, $do_wh64_w	# U :
+
+	nop
+	nop
+	nop
+	beq	$3, no_quad_w	# U : Might have finished already
+
+.align 4
+	/*
+	 * Simple loop for trailing quadwords, or for small amounts
+	 * of data (where we can't use an unrolled loop and wh64)
+	 */
+loop_w:
+	stq $17,0($5)		# L :
+	subq $3,1,$3		# E : Decrement number quads left
+	addq $5,8,$5		# E : Inc address
+	bne $3,loop_w		# U : more?
+
+no_quad_w:
+	/*
+	 * Write 0..7 trailing bytes.
+	 */
+	nop			# E :
+	beq $18,end_w		# U : All done?
+	ldq $7,0($5)		# L :
+	mskqh $7,$6,$2		# U : Mask final quad
+
+	insqh $17,$6,$4		# U : New bits
+	bis $2,$4,$1		# E : Put it all together
+	stq $1,0($5)		# L : And back to memory
+	ret $31,($26),1		# L0 :
+
+within_quad_w:
+	ldq_u $1,0($16)		# L :
+	insql $17,$16,$2	# U : New bits
+	mskql $1,$16,$4		# U : Clear old
+	bis $2,$4,$2		# E : New result
+
+	mskql $2,$6,$4		# U :
+	mskqh $1,$6,$2		# U :
+	bis $2,$4,$1		# E :
+	stq_u $1,0($16)		# L :
+
+end_w:
+	nop
+	nop
+	nop
+	ret $31,($26),1		# L0 :
+
+	.end __memsetw
+
+memset = ___memset
+__memset = ___memset

arch/alpha/lib/ev6-stxcpy.S

@@ -0,0 +1,321 @@
+/*
+ * arch/alpha/lib/ev6-stxcpy.S
+ * 21264 version contributed by Rick Gorton <rick.gorton@alpha-processor.com>
+ *
+ * Copy a null-terminated string from SRC to DST.
+ *
+ * This is an internal routine used by strcpy, stpcpy, and strcat.
+ * As such, it uses special linkage conventions to make implementation
+ * of these public functions more efficient.
+ *
+ * On input:
+ *	t9 = return address
+ *	a0 = DST
+ *	a1 = SRC
+ *
+ * On output:
+ *	t12 = bitmask (with one bit set) indicating the last byte written
+ *	a0  = unaligned address of the last *word* written
+ *
+ * Furthermore, v0, a3-a5, t11, and t12 are untouched.
+ *
+ * Much of the information about 21264 scheduling/coding comes from:
+ *	Compiler Writer's Guide for the Alpha 21264
+ *	abbreviated as 'CWG' in other comments here
+ *	ftp.digital.com/pub/Digital/info/semiconductor/literature/dsc-library.html
+ * Scheduling notation:
+ *	E	- either cluster
+ *	U	- upper subcluster; U0 - subcluster U0; U1 - subcluster U1
+ *	L	- lower subcluster; L0 - subcluster L0; L1 - subcluster L1
+ * Try not to change the actual algorithm if possible for consistency.
+ */
+
+#include <asm/regdef.h>
+
+	.set noat
+	.set noreorder
+
+	.text
+
+/* There is a problem with either gdb (as of 4.16) or gas (as of 2.7) that
+   doesn't like putting the entry point for a procedure somewhere in the
+   middle of the procedure descriptor.  Work around this by putting the
+   aligned copy in its own procedure descriptor */
+
+
+	.ent stxcpy_aligned
+	.align 4
+stxcpy_aligned:
+	.frame sp, 0, t9
+	.prologue 0
+
+	/* On entry to this basic block:
+	   t0 == the first destination word for masking back in
+	   t1 == the first source word.  */
+
+	/* Create the 1st output word and detect 0's in the 1st input word.  */
+	lda	t2, -1		# E : build a mask against false zero
+	mskqh	t2, a1, t2	# U :   detection in the src word (stall)
+	mskqh	t1, a1, t3	# U :
+	ornot	t1, t2, t2	# E : (stall)
+
+	mskql	t0, a1, t0	# U : assemble the first output word
+	cmpbge	zero, t2, t8	# E : bits set iff null found
+	or	t0, t3, t1	# E : (stall)
+	bne	t8, $a_eos	# U : (stall)
+
+	/* On entry to this basic block:
+	   t0 == the first destination word for masking back in
+	   t1 == a source word not containing a null.  */
+	/* Nops here to separate store quads from load quads */
+
+$a_loop:
+	stq_u	t1, 0(a0)	# L :
+	addq	a0, 8, a0	# E :
+	nop
+	nop
+
+	ldq_u	t1, 0(a1)	# L : Latency=3
+	addq	a1, 8, a1	# E :
+	cmpbge	zero, t1, t8	# E : (3 cycle stall)
+	beq	t8, $a_loop	# U : (stall for t8)
+
+	/* Take care of the final (partial) word store.
+	   On entry to this basic block we have:
+	   t1 == the source word containing the null
+	   t8 == the cmpbge mask that found it.  */
+$a_eos:
+	negq	t8, t6		# E : find low bit set
+	and	t8, t6, t12	# E : (stall)
+	/* For the sake of the cache, don't read a destination word
+	   if we're not going to need it.  */
+	and	t12, 0x80, t6	# E : (stall)
+	bne	t6, 1f		# U : (stall)
+
+	/* We're doing a partial word store and so need to combine
+	   our source and original destination words.  */
+	ldq_u	t0, 0(a0)	# L : Latency=3
+	subq	t12, 1, t6	# E :
+	zapnot	t1, t6, t1	# U : clear src bytes >= null (stall)
+	or	t12, t6, t8	# E : (stall)
+
+	zap	t0, t8, t0	# E : clear dst bytes <= null
+	or	t0, t1, t1	# E : (stall)
+	nop
+	nop
+
+1:	stq_u	t1, 0(a0)	# L :
+	ret	(t9)		# L0 : Latency=3
+	nop
+	nop
+
+	.end stxcpy_aligned
+
+	.align 4
+	.ent __stxcpy
+	.globl __stxcpy
+__stxcpy:
+	.frame sp, 0, t9
+	.prologue 0
+
+	/* Are source and destination co-aligned?  */
+	xor	a0, a1, t0	# E :
+	unop			# E :
+	and	t0, 7, t0	# E : (stall)
+	bne	t0, $unaligned	# U : (stall)
+
+	/* We are co-aligned; take care of a partial first word.  */
+	ldq_u	t1, 0(a1)		# L : load first src word
+	and	a0, 7, t0		# E : take care not to load a word ...
+	addq	a1, 8, a1		# E :
+	beq	t0, stxcpy_aligned	# U : ... if we wont need it (stall)
+
+	ldq_u	t0, 0(a0)	# L :
+	br	stxcpy_aligned	# L0 : Latency=3
+	nop
+	nop
+
+
+/* The source and destination are not co-aligned.  Align the destination
+   and cope.  We have to be very careful about not reading too much and
+   causing a SEGV.  */
+
+	.align 4
+$u_head:
+	/* We know just enough now to be able to assemble the first
+	   full source word.  We can still find a zero at the end of it
+	   that prevents us from outputting the whole thing.
+
+	   On entry to this basic block:
+	   t0 == the first dest word, for masking back in, if needed else 0
+	   t1 == the low bits of the first source word
+	   t6 == bytemask that is -1 in dest word bytes */
+
+	ldq_u	t2, 8(a1)	# L :
+	addq	a1, 8, a1	# E :
+	extql	t1, a1, t1	# U : (stall on a1)
+	extqh	t2, a1, t4	# U : (stall on a1)
+
+	mskql	t0, a0, t0	# U :
+	or	t1, t4, t1	# E :
+	mskqh	t1, a0, t1	# U : (stall on t1)
+	or	t0, t1, t1	# E : (stall on t1)
+
+	or	t1, t6, t6	# E :
+	cmpbge	zero, t6, t8	# E : (stall)
+	lda	t6, -1		# E : for masking just below
+	bne	t8, $u_final	# U : (stall)
+
+	mskql	t6, a1, t6		# U : mask out the bits we have
+	or	t6, t2, t2		# E :   already extracted before (stall)
+	cmpbge	zero, t2, t8		# E :   testing eos (stall)
+	bne	t8, $u_late_head_exit	# U : (stall)
+
+	/* Finally, we've got all the stupid leading edge cases taken care
+	   of and we can set up to enter the main loop.  */
+
+	stq_u	t1, 0(a0)	# L : store first output word
+	addq	a0, 8, a0	# E :
+	extql	t2, a1, t0	# U : position ho-bits of lo word
+	ldq_u	t2, 8(a1)	# U : read next high-order source word
+
+	addq	a1, 8, a1	# E :
+	cmpbge	zero, t2, t8	# E : (stall for t2)
+	nop			# E :
+	bne	t8, $u_eos	# U : (stall)
+
+	/* Unaligned copy main loop.  In order to avoid reading too much,
+	   the loop is structured to detect zeros in aligned source words.
+	   This has, unfortunately, effectively pulled half of a loop
+	   iteration out into the head and half into the tail, but it does
+	   prevent nastiness from accumulating in the very thing we want
+	   to run as fast as possible.
+
+	   On entry to this basic block:
+	   t0 == the shifted high-order bits from the previous source word
+	   t2 == the unshifted current source word
+
+	   We further know that t2 does not contain a null terminator.  */
+
+	.align 3
+$u_loop:
+	extqh	t2, a1, t1	# U : extract high bits for current word
+	addq	a1, 8, a1	# E : (stall)
+	extql	t2, a1, t3	# U : extract low bits for next time (stall)
+	addq	a0, 8, a0	# E :
+
+	or	t0, t1, t1	# E : current dst word now complete
+	ldq_u	t2, 0(a1)	# L : Latency=3 load high word for next time
+	stq_u	t1, -8(a0)	# L : save the current word (stall)
+	mov	t3, t0		# E :
+
+	cmpbge	zero, t2, t8	# E : test new word for eos
+	beq	t8, $u_loop	# U : (stall)
+	nop
+	nop
+
+	/* We've found a zero somewhere in the source word we just read.
+	   If it resides in the lower half, we have one (probably partial)
+	   word to write out, and if it resides in the upper half, we
+	   have one full and one partial word left to write out.
+
+	   On entry to this basic block:
+	   t0 == the shifted high-order bits from the previous source word
+	   t2 == the unshifted current source word.  */
+$u_eos:
+	extqh	t2, a1, t1	# U :
+	or	t0, t1, t1	# E : first (partial) source word complete (stall)
+	cmpbge	zero, t1, t8	# E : is the null in this first bit? (stall)
+	bne	t8, $u_final	# U : (stall)
+
+$u_late_head_exit:
+	stq_u	t1, 0(a0)	# L : the null was in the high-order bits
+	addq	a0, 8, a0	# E :
+	extql	t2, a1, t1	# U :
+	cmpbge	zero, t1, t8	# E : (stall)
+
+	/* Take care of a final (probably partial) result word.
+	   On entry to this basic block:
+	   t1 == assembled source word
+	   t8 == cmpbge mask that found the null.  */
+$u_final:
+	negq	t8, t6		# E : isolate low bit set
+	and	t6, t8, t12	# E : (stall)
+	and	t12, 0x80, t6	# E : avoid dest word load if we can (stall)
+	bne	t6, 1f		# U : (stall)
+
+	ldq_u	t0, 0(a0)	# E :
+	subq	t12, 1, t6	# E :
+	or	t6, t12, t8	# E : (stall)
+	zapnot	t1, t6, t1	# U : kill source bytes >= null (stall)
+
+	zap	t0, t8, t0	# U : kill dest bytes <= null (2 cycle data stall)
+	or	t0, t1, t1	# E : (stall)
+	nop
+	nop
+
+1:	stq_u	t1, 0(a0)	# L :
+	ret	(t9)		# L0 : Latency=3
+	nop
+	nop
+
+	/* Unaligned copy entry point.  */
+	.align 4
+$unaligned:
+
+	ldq_u	t1, 0(a1)	# L : load first source word
+	and	a0, 7, t4	# E : find dest misalignment
+	and	a1, 7, t5	# E : find src misalignment
+	/* Conditionally load the first destination word and a bytemask
+	   with 0xff indicating that the destination byte is sacrosanct.  */
+	mov	zero, t0	# E :
+
+	mov	zero, t6	# E :
+	beq	t4, 1f		# U :
+	ldq_u	t0, 0(a0)	# L :
+	lda	t6, -1		# E :
+
+	mskql	t6, a0, t6	# U :
+	nop
+	nop
+	nop
+1:
+	subq	a1, t4, a1	# E : sub dest misalignment from src addr
+	/* If source misalignment is larger than dest misalignment, we need
+	   extra startup checks to avoid SEGV.  */
+	cmplt	t4, t5, t12	# E :
+	beq	t12, $u_head	# U :
+	lda	t2, -1		# E : mask out leading garbage in source
+
+	mskqh	t2, t5, t2	# U :
+	ornot	t1, t2, t3	# E : (stall)
+	cmpbge	zero, t3, t8	# E : is there a zero? (stall)
+	beq	t8, $u_head	# U : (stall)
+
+	/* At this point we've found a zero in the first partial word of
+	   the source.  We need to isolate the valid source data and mask
+	   it into the original destination data.  (Incidentally, we know
+	   that we'll need at least one byte of that original dest word.) */
+
+	ldq_u	t0, 0(a0)	# L :
+	negq	t8, t6		# E : build bitmask of bytes <= zero
+	and	t6, t8, t12	# E : (stall)
+	and	a1, 7, t5	# E :
+
+	subq	t12, 1, t6	# E :
+	or	t6, t12, t8	# E : (stall)
+	srl	t12, t5, t12	# U : adjust final null return value
+	zapnot	t2, t8, t2	# U : prepare source word; mirror changes (stall)
+
+	and	t1, t2, t1	# E : to source validity mask
+	extql	t2, a1, t2	# U :
+	extql	t1, a1, t1	# U : (stall)
+	andnot	t0, t2, t0	# .. e1 : zero place for source to reside (stall)
+
+	or	t0, t1, t1	# e1    : and put it there
+	stq_u	t1, 0(a0)	# .. e0 : (stall)
+	ret	(t9)		# e1    :
+	nop
+
+	.end __stxcpy
+

arch/alpha/lib/ev6-stxncpy.S

@@ -0,0 +1,397 @@
+/*
+ * arch/alpha/lib/ev6-stxncpy.S
+ * 21264 version contributed by Rick Gorton <rick.gorton@api-networks.com>
+ *
+ * Copy no more than COUNT bytes of the null-terminated string from
+ * SRC to DST.
+ *
+ * This is an internal routine used by strncpy, stpncpy, and strncat.
+ * As such, it uses special linkage conventions to make implementation
+ * of these public functions more efficient.
+ *
+ * On input:
+ *	t9 = return address
+ *	a0 = DST
+ *	a1 = SRC
+ *	a2 = COUNT
+ *
+ * Furthermore, COUNT may not be zero.
+ *
+ * On output:
+ *	t0  = last word written
+ *	t10 = bitmask (with one bit set) indicating the byte position of
+ *	      the end of the range specified by COUNT
+ *	t12 = bitmask (with one bit set) indicating the last byte written
+ *	a0  = unaligned address of the last *word* written
+ *	a2  = the number of full words left in COUNT
+ *
+ * Furthermore, v0, a3-a5, t11, and $at are untouched.
+ *
+ * Much of the information about 21264 scheduling/coding comes from:
+ *	Compiler Writer's Guide for the Alpha 21264
+ *	abbreviated as 'CWG' in other comments here
+ *	ftp.digital.com/pub/Digital/info/semiconductor/literature/dsc-library.html
+ * Scheduling notation:
+ *	E	- either cluster
+ *	U	- upper subcluster; U0 - subcluster U0; U1 - subcluster U1
+ *	L	- lower subcluster; L0 - subcluster L0; L1 - subcluster L1
+ * Try not to change the actual algorithm if possible for consistency.
+ */
+
+#include <asm/regdef.h>
+
+	.set noat
+	.set noreorder
+
+	.text
+
+/* There is a problem with either gdb (as of 4.16) or gas (as of 2.7) that
+   doesn't like putting the entry point for a procedure somewhere in the
+   middle of the procedure descriptor.  Work around this by putting the
+   aligned copy in its own procedure descriptor */
+
+
+	.ent stxncpy_aligned
+	.align 4
+stxncpy_aligned:
+	.frame sp, 0, t9, 0
+	.prologue 0
+
+	/* On entry to this basic block:
+	   t0 == the first destination word for masking back in
+	   t1 == the first source word.  */
+
+	/* Create the 1st output word and detect 0's in the 1st input word.  */
+	lda	t2, -1		# E : build a mask against false zero
+	mskqh	t2, a1, t2	# U :   detection in the src word (stall)
+	mskqh	t1, a1, t3	# U :
+	ornot	t1, t2, t2	# E : (stall)
+
+	mskql	t0, a1, t0	# U : assemble the first output word
+	cmpbge	zero, t2, t8	# E : bits set iff null found
+	or	t0, t3, t0	# E : (stall)
+	beq	a2, $a_eoc	# U :
+
+	bne	t8, $a_eos	# U :
+	nop
+	nop
+	nop
+
+	/* On entry to this basic block:
+	   t0 == a source word not containing a null.  */
+
+	/*
+	 * nops here to:
+	 *	separate store quads from load quads
+	 *	limit of 1 bcond/quad to permit training
+	 */
+$a_loop:
+	stq_u	t0, 0(a0)	# L :
+	addq	a0, 8, a0	# E :
+	subq	a2, 1, a2	# E :
+	nop
+
+	ldq_u	t0, 0(a1)	# L :
+	addq	a1, 8, a1	# E :
+	cmpbge	zero, t0, t8	# E :
+	beq	a2, $a_eoc      # U :
+
+	beq	t8, $a_loop	# U :
+	nop
+	nop
+	nop
+
+	/* Take care of the final (partial) word store.  At this point
+	   the end-of-count bit is set in t8 iff it applies.
+
+	   On entry to this basic block we have:
+	   t0 == the source word containing the null
+	   t8 == the cmpbge mask that found it.  */
+
+$a_eos:
+	negq	t8, t12		# E : find low bit set
+	and	t8, t12, t12	# E : (stall)
+	/* For the sake of the cache, don't read a destination word
+	   if we're not going to need it.  */
+	and	t12, 0x80, t6	# E : (stall)
+	bne	t6, 1f		# U : (stall)
+
+	/* We're doing a partial word store and so need to combine
+	   our source and original destination words.  */
+	ldq_u	t1, 0(a0)	# L :
+	subq	t12, 1, t6	# E :
+	or	t12, t6, t8	# E : (stall)
+	zapnot	t0, t8, t0	# U : clear src bytes > null (stall)
+
+	zap	t1, t8, t1	# .. e1 : clear dst bytes <= null
+	or	t0, t1, t0	# e1    : (stall)
+	nop
+	nop
+
+1:	stq_u	t0, 0(a0)	# L :
+	ret	(t9)		# L0 : Latency=3
+	nop
+	nop
+
+	/* Add the end-of-count bit to the eos detection bitmask.  */
+$a_eoc:
+	or	t10, t8, t8	# E :
+	br	$a_eos		# L0 : Latency=3
+	nop
+	nop
+
+	.end stxncpy_aligned
+
+	.align 4
+	.ent __stxncpy
+	.globl __stxncpy
+__stxncpy:
+	.frame sp, 0, t9, 0
+	.prologue 0
+
+	/* Are source and destination co-aligned?  */
+	xor	a0, a1, t1	# E :
+	and	a0, 7, t0	# E : find dest misalignment
+	and	t1, 7, t1	# E : (stall)
+	addq	a2, t0, a2	# E : bias count by dest misalignment (stall)
+
+	subq	a2, 1, a2	# E :
+	and	a2, 7, t2	# E : (stall)
+	srl	a2, 3, a2	# U : a2 = loop counter = (count - 1)/8 (stall)
+	addq	zero, 1, t10	# E :
+
+	sll	t10, t2, t10	# U : t10 = bitmask of last count byte
+	bne	t1, $unaligned	# U :
+	/* We are co-aligned; take care of a partial first word.  */
+	ldq_u	t1, 0(a1)	# L : load first src word
+	addq	a1, 8, a1	# E :
+
+	beq	t0, stxncpy_aligned     # U : avoid loading dest word if not needed
+	ldq_u	t0, 0(a0)	# L :
+	nop
+	nop
+
+	br	stxncpy_aligned	# .. e1 :
+	nop
+	nop
+	nop
+
+
+
+/* The source and destination are not co-aligned.  Align the destination
+   and cope.  We have to be very careful about not reading too much and
+   causing a SEGV.  */
+
+	.align 4
+$u_head:
+	/* We know just enough now to be able to assemble the first
+	   full source word.  We can still find a zero at the end of it
+	   that prevents us from outputting the whole thing.
+
+	   On entry to this basic block:
+	   t0 == the first dest word, unmasked
+	   t1 == the shifted low bits of the first source word
+	   t6 == bytemask that is -1 in dest word bytes */
+
+	ldq_u	t2, 8(a1)	# L : Latency=3 load second src word
+	addq	a1, 8, a1	# E :
+	mskql	t0, a0, t0	# U : mask trailing garbage in dst
+	extqh	t2, a1, t4	# U : (3 cycle stall on t2)
+
+	or	t1, t4, t1	# E : first aligned src word complete (stall)
+	mskqh	t1, a0, t1	# U : mask leading garbage in src (stall)
+	or	t0, t1, t0	# E : first output word complete (stall)
+	or	t0, t6, t6	# E : mask original data for zero test (stall)
+
+	cmpbge	zero, t6, t8	# E :
+	beq	a2, $u_eocfin	# U :
+	lda	t6, -1		# E :
+	nop
+
+	bne	t8, $u_final	# U :
+	mskql	t6, a1, t6	# U : mask out bits already seen
+	stq_u	t0, 0(a0)	# L : store first output word
+	or      t6, t2, t2	# E : (stall)
+
+	cmpbge	zero, t2, t8	# E : find nulls in second partial
+	addq	a0, 8, a0	# E :
+	subq	a2, 1, a2	# E :
+	bne	t8, $u_late_head_exit	# U :
+
+	/* Finally, we've got all the stupid leading edge cases taken care
+	   of and we can set up to enter the main loop.  */
+	extql	t2, a1, t1	# U : position hi-bits of lo word
+	beq	a2, $u_eoc	# U :
+	ldq_u	t2, 8(a1)	# L : read next high-order source word
+	addq	a1, 8, a1	# E :
+
+	extqh	t2, a1, t0	# U : position lo-bits of hi word (stall)
+	cmpbge	zero, t2, t8	# E :
+	nop
+	bne	t8, $u_eos	# U :
+
+	/* Unaligned copy main loop.  In order to avoid reading too much,
+	   the loop is structured to detect zeros in aligned source words.
+	   This has, unfortunately, effectively pulled half of a loop
+	   iteration out into the head and half into the tail, but it does
+	   prevent nastiness from accumulating in the very thing we want
+	   to run as fast as possible.
+
+	   On entry to this basic block:
+	   t0 == the shifted low-order bits from the current source word
+	   t1 == the shifted high-order bits from the previous source word
+	   t2 == the unshifted current source word
+
+	   We further know that t2 does not contain a null terminator.  */
+
+	.align 4
+$u_loop:
+	or	t0, t1, t0	# E : current dst word now complete
+	subq	a2, 1, a2	# E : decrement word count
+	extql	t2, a1, t1	# U : extract low bits for next time
+	addq	a0, 8, a0	# E :
+
+	stq_u	t0, -8(a0)	# U : save the current word
+	beq	a2, $u_eoc	# U :
+	ldq_u	t2, 8(a1)	# U : Latency=3 load high word for next time
+	addq	a1, 8, a1	# E :
+
+	extqh	t2, a1, t0	# U : extract low bits (2 cycle stall)
+	cmpbge	zero, t2, t8	# E : test new word for eos
+	nop
+	beq	t8, $u_loop	# U :
+
+	/* We've found a zero somewhere in the source word we just read.
+	   If it resides in the lower half, we have one (probably partial)
+	   word to write out, and if it resides in the upper half, we
+	   have one full and one partial word left to write out.
+
+	   On entry to this basic block:
+	   t0 == the shifted low-order bits from the current source word
+	   t1 == the shifted high-order bits from the previous source word
+	   t2 == the unshifted current source word.  */
+$u_eos:
+	or	t0, t1, t0	# E : first (partial) source word complete
+	nop
+	cmpbge	zero, t0, t8	# E : is the null in this first bit? (stall)
+	bne	t8, $u_final	# U : (stall)
+
+	stq_u	t0, 0(a0)	# L : the null was in the high-order bits
+	addq	a0, 8, a0	# E :
+	subq	a2, 1, a2	# E :
+	nop
+
+$u_late_head_exit:
+	extql	t2, a1, t0	# U :
+	cmpbge	zero, t0, t8	# E :
+	or	t8, t10, t6	# E : (stall)
+	cmoveq	a2, t6, t8	# E : Latency=2, extra map slot (stall)
+
+	/* Take care of a final (probably partial) result word.
+	   On entry to this basic block:
+	   t0 == assembled source word
+	   t8 == cmpbge mask that found the null.  */
+$u_final:
+	negq	t8, t6		# E : isolate low bit set
+	and	t6, t8, t12	# E : (stall)
+	and	t12, 0x80, t6	# E : avoid dest word load if we can (stall)
+	bne	t6, 1f		# U : (stall)
+
+	ldq_u	t1, 0(a0)	# L :
+	subq	t12, 1, t6	# E :
+	or	t6, t12, t8	# E : (stall)
+	zapnot	t0, t8, t0	# U : kill source bytes > null
+
+	zap	t1, t8, t1	# U : kill dest bytes <= null
+	or	t0, t1, t0	# E : (stall)
+	nop
+	nop
+
+1:	stq_u	t0, 0(a0)	# L :
+	ret	(t9)		# L0 : Latency=3
+
+	  /* Got to end-of-count before end of string.  
+	     On entry to this basic block:
+	     t1 == the shifted high-order bits from the previous source word  */
+$u_eoc:
+	and	a1, 7, t6	# E : avoid final load if possible
+	sll	t10, t6, t6	# U : (stall)
+	and	t6, 0xff, t6	# E : (stall)
+	bne	t6, 1f		# U : (stall)
+
+	ldq_u	t2, 8(a1)	# L : load final src word
+	nop
+	extqh	t2, a1, t0	# U : extract low bits for last word (stall)
+	or	t1, t0, t1	# E : (stall)
+
+1:	cmpbge	zero, t1, t8	# E :
+	mov	t1, t0		# E :
+
+$u_eocfin:			# end-of-count, final word
+	or	t10, t8, t8	# E :
+	br	$u_final	# L0 : Latency=3
+
+	/* Unaligned copy entry point.  */
+	.align 4
+$unaligned:
+
+	ldq_u	t1, 0(a1)	# L : load first source word
+	and	a0, 7, t4	# E : find dest misalignment
+	and	a1, 7, t5	# E : find src misalignment
+	/* Conditionally load the first destination word and a bytemask
+	   with 0xff indicating that the destination byte is sacrosanct.  */
+	mov	zero, t0	# E :
+
+	mov	zero, t6	# E :
+	beq	t4, 1f		# U :
+	ldq_u	t0, 0(a0)	# L :
+	lda	t6, -1		# E :
+
+	mskql	t6, a0, t6	# U :
+	nop
+	nop
+	subq	a1, t4, a1	# E : sub dest misalignment from src addr
+
+	/* If source misalignment is larger than dest misalignment, we need
+	   extra startup checks to avoid SEGV.  */
+
+1:	cmplt	t4, t5, t12	# E :
+	extql	t1, a1, t1	# U : shift src into place
+	lda	t2, -1		# E : for creating masks later
+	beq	t12, $u_head	# U : (stall)
+
+	extql	t2, a1, t2	# U :
+	cmpbge	zero, t1, t8	# E : is there a zero?
+	andnot	t2, t6, t2	# E : dest mask for a single word copy
+	or	t8, t10, t5	# E : test for end-of-count too
+
+	cmpbge	zero, t2, t3	# E :
+	cmoveq	a2, t5, t8	# E : Latency=2, extra map slot
+	nop			# E : keep with cmoveq
+	andnot	t8, t3, t8	# E : (stall)
+
+	beq	t8, $u_head	# U :
+	/* At this point we've found a zero in the first partial word of
+	   the source.  We need to isolate the valid source data and mask
+	   it into the original destination data.  (Incidentally, we know
+	   that we'll need at least one byte of that original dest word.) */
+	ldq_u	t0, 0(a0)	# L :
+	negq	t8, t6		# E : build bitmask of bytes <= zero
+	mskqh	t1, t4, t1	# U :
+
+	and	t6, t8, t12	# E :
+	subq	t12, 1, t6	# E : (stall)
+	or	t6, t12, t8	# E : (stall)
+	zapnot	t2, t8, t2	# U : prepare source word; mirror changes (stall)
+
+	zapnot	t1, t8, t1	# U : to source validity mask
+	andnot	t0, t2, t0	# E : zero place for source to reside
+	or	t0, t1, t0	# E : and put it there (stall both t0, t1)
+	stq_u	t0, 0(a0)	# L : (stall)
+
+	ret	(t9)		# L0 : Latency=3
+	nop
+	nop
+	nop
+
+	.end __stxncpy

arch/alpha/lib/ev67-strcat.S

@@ -0,0 +1,54 @@
+/*
+ * arch/alpha/lib/ev67-strcat.S
+ * 21264 version contributed by Rick Gorton <rick.gorton@alpha-processor.com>
+ *
+ * Append a null-terminated string from SRC to DST.
+ *
+ * Much of the information about 21264 scheduling/coding comes from:
+ *	Compiler Writer's Guide for the Alpha 21264
+ *	abbreviated as 'CWG' in other comments here
+ *	ftp.digital.com/pub/Digital/info/semiconductor/literature/dsc-library.html
+ * Scheduling notation:
+ *	E	- either cluster
+ *	U	- upper subcluster; U0 - subcluster U0; U1 - subcluster U1
+ *	L	- lower subcluster; L0 - subcluster L0; L1 - subcluster L1
+ * Try not to change the actual algorithm if possible for consistency.
+ * Commentary: It seems bogus to walk the input string twice - once
+ * to determine the length, and then again while doing the copy.
+ * A significant (future) enhancement would be to only read the input
+ * string once.
+ */
+
+
+	.text
+
+	.align 4
+	.globl strcat
+	.ent strcat
+strcat:
+	.frame $30, 0, $26
+	.prologue 0
+
+	mov	$16, $0		# E : set up return value
+	/* Find the end of the string.  */
+	ldq_u   $1, 0($16)	# L : load first quadword (a0 may be misaligned)
+	lda     $2, -1		# E :
+	insqh   $2, $16, $2	# U :
+
+	andnot  $16, 7, $16	# E :
+	or      $2, $1, $1	# E :
+	cmpbge  $31, $1, $2	# E : bits set iff byte == 0
+	bne     $2, $found	# U :
+
+$loop:	ldq     $1, 8($16)	# L :
+	addq    $16, 8, $16	# E :
+	cmpbge  $31, $1, $2	# E :
+	beq     $2, $loop	# U :
+
+$found:	cttz	$2, $3		# U0 :
+	addq	$16, $3, $16	# E :
+	/* Now do the append.  */
+	mov	$26, $23	# E :
+	br	__stxcpy	# L0 :
+
+	.end strcat

arch/alpha/lib/ev67-strchr.S

@@ -0,0 +1,88 @@
+/*
+ * arch/alpha/lib/ev67-strchr.S
+ * 21264 version contributed by Rick Gorton <rick.gorton@alpha-processor.com>
+ *
+ * Return the address of a given character within a null-terminated
+ * string, or null if it is not found.
+ *
+ * Much of the information about 21264 scheduling/coding comes from:
+ *	Compiler Writer's Guide for the Alpha 21264
+ *	abbreviated as 'CWG' in other comments here
+ *	ftp.digital.com/pub/Digital/info/semiconductor/literature/dsc-library.html
+ * Scheduling notation:
+ *	E	- either cluster
+ *	U	- upper subcluster; U0 - subcluster U0; U1 - subcluster U1
+ *	L	- lower subcluster; L0 - subcluster L0; L1 - subcluster L1
+ * Try not to change the actual algorithm if possible for consistency.
+ */
+
+#include <asm/regdef.h>
+
+	.set noreorder
+	.set noat
+
+	.align 4
+	.globl strchr
+	.ent strchr
+strchr:
+	.frame sp, 0, ra
+	.prologue 0
+
+	ldq_u   t0, 0(a0)	# L : load first quadword Latency=3
+	and	a1, 0xff, t3	# E : 00000000000000ch
+	insbl	a1, 1, t5	# U : 000000000000ch00
+	insbl	a1, 7, a2	# U : ch00000000000000
+
+	insbl	t3, 6, a3	# U : 00ch000000000000
+	or	t5, t3, a1	# E : 000000000000chch
+	andnot  a0, 7, v0	# E : align our loop pointer
+	lda	t4, -1		# E : build garbage mask
+
+	mskqh	t4, a0, t4	# U : only want relevant part of first quad
+	or	a2, a3, a2	# E : chch000000000000
+	inswl	a1, 2, t5	# E : 00000000chch0000
+	inswl	a1, 4, a3	# E : 0000chch00000000
+
+	or	a1, a2, a1	# E : chch00000000chch
+	or	a3, t5, t5	# E : 0000chchchch0000
+	cmpbge  zero, t0, t2	# E : bits set iff byte == zero
+	cmpbge	zero, t4, t4	# E : bits set iff byte is garbage
+
+	/* This quad is _very_ serialized.  Lots of stalling happens */
+	or	t5, a1, a1	# E : chchchchchchchch
+	xor	t0, a1, t1	# E : make bytes == c zero
+	cmpbge  zero, t1, t3	# E : bits set iff byte == c
+	or	t2, t3, t0	# E : bits set iff char match or zero match
+
+	andnot	t0, t4, t0	# E : clear garbage bits
+	cttz	t0, a2		# U0 : speculative (in case we get a match)
+	nop			# E :
+	bne	t0, $found	# U :
+
+	/*
+	 * Yuk.  This loop is going to stall like crazy waiting for the
+	 * data to be loaded.  Not much can be done about it unless it's
+	 * unrolled multiple times - is that safe to do in kernel space?
+	 * Or would exception handling recovery code do the trick here?
+	 */
+$loop:	ldq	t0, 8(v0)	# L : Latency=3
+	addq	v0, 8, v0	# E :
+	xor	t0, a1, t1	# E :
+	cmpbge	zero, t0, t2	# E : bits set iff byte == 0
+
+	cmpbge	zero, t1, t3	# E : bits set iff byte == c
+	or	t2, t3, t0	# E :
+	cttz	t3, a2		# U0 : speculative (in case we get a match)
+	beq	t0, $loop	# U :
+
+$found:	negq    t0, t1		# E : clear all but least set bit
+	and     t0, t1, t0	# E :
+	and	t0, t3, t1	# E : bit set iff byte was the char
+	addq	v0, a2, v0	# E : Add in the bit number from above
+
+	cmoveq	t1, $31, v0	# E : Two mapping slots, latency = 2
+	nop
+	nop
+	ret			# L0 :
+
+	.end strchr

arch/alpha/lib/ev67-strlen.S

@@ -0,0 +1,49 @@
+/*
+ * arch/alpha/lib/ev67-strlen.S
+ * 21264 version by Rick Gorton <rick.gorton@alpha-processor.com>
+ *
+ * Finds length of a 0-terminated string.  Optimized for the
+ * Alpha architecture:
+ *
+ *	- memory accessed as aligned quadwords only
+ *	- uses bcmpge to compare 8 bytes in parallel
+ *
+ * Much of the information about 21264 scheduling/coding comes from:
+ *	Compiler Writer's Guide for the Alpha 21264
+ *	abbreviated as 'CWG' in other comments here
+ *	ftp.digital.com/pub/Digital/info/semiconductor/literature/dsc-library.html
+ * Scheduling notation:
+ *	E	- either cluster
+ *	U	- upper subcluster; U0 - subcluster U0; U1 - subcluster U1
+ *	L	- lower subcluster; L0 - subcluster L0; L1 - subcluster L1
+ */
+
+	.set noreorder
+	.set noat
+
+	.globl	strlen
+	.ent	strlen
+	.align 4
+strlen:
+	ldq_u	$1, 0($16)	# L : load first quadword ($16  may be misaligned)
+	lda	$2, -1($31)	# E :
+	insqh	$2, $16, $2	# U :
+	andnot	$16, 7, $0	# E :
+
+	or	$2, $1, $1	# E :
+	cmpbge	$31, $1, $2	# E : $2  <- bitmask: bit i == 1 <==> i-th byte == 0
+	nop			# E :
+	bne	$2, $found	# U :
+
+$loop:	ldq	$1, 8($0)	# L :
+	addq	$0, 8, $0	# E : addr += 8
+	cmpbge	$31, $1, $2	# E :
+	beq	$2, $loop	# U :
+
+$found:
+	cttz	$2, $3		# U0 :
+	addq	$0, $3, $0	# E :
+	subq	$0, $16, $0	# E :
+	ret	$31, ($26)	# L0 :
+
+	.end	strlen

arch/alpha/lib/ev67-strncat.S

@@ -0,0 +1,94 @@
+/*
+ * arch/alpha/lib/ev67-strncat.S
+ * 21264 version contributed by Rick Gorton <rick.gorton@api-networks.com>
+ *
+ * Append no more than COUNT characters from the null-terminated string SRC
+ * to the null-terminated string DST.  Always null-terminate the new DST.
+ *
+ * This differs slightly from the semantics in libc in that we never write
+ * past count, whereas libc may write to count+1.  This follows the generic
+ * implementation in lib/string.c and is, IMHO, more sensible.
+ *
+ * Much of the information about 21264 scheduling/coding comes from:
+ *	Compiler Writer's Guide for the Alpha 21264
+ *	abbreviated as 'CWG' in other comments here
+ *	ftp.digital.com/pub/Digital/info/semiconductor/literature/dsc-library.html
+ * Scheduling notation:
+ *	E	- either cluster
+ *	U	- upper subcluster; U0 - subcluster U0; U1 - subcluster U1
+ *	L	- lower subcluster; L0 - subcluster L0; L1 - subcluster L1
+ * Try not to change the actual algorithm if possible for consistency.
+ */
+
+
+	.text
+
+	.align 4
+	.globl strncat
+	.ent strncat
+strncat:
+	.frame $30, 0, $26
+	.prologue 0
+
+	mov	$16, $0		# set up return value
+	beq	$18, $zerocount	# U :
+	/* Find the end of the string.  */
+	ldq_u   $1, 0($16)	# L : load first quadword ($16 may be misaligned)
+	lda     $2, -1($31)	# E :
+
+	insqh   $2, $0, $2	# U :
+	andnot  $16, 7, $16	# E :
+	nop			# E :
+	or      $2, $1, $1	# E :
+
+	nop			# E :
+	nop			# E :
+	cmpbge  $31, $1, $2	# E : bits set iff byte == 0
+	bne     $2, $found	# U :
+
+$loop:	ldq     $1, 8($16)	# L :
+	addq    $16, 8, $16	# E :
+	cmpbge  $31, $1, $2	# E :
+	beq     $2, $loop	# U :
+
+$found:	cttz	$2, $3		# U0 :
+	addq	$16, $3, $16	# E :
+	nop			# E :
+	bsr	$23, __stxncpy	# L0 :/* Now do the append.  */
+
+	/* Worry about the null termination.  */
+
+	zapnot	$1, $27, $2	# U : was last byte a null?
+	cmplt	$27, $24, $5	# E : did we fill the buffer completely?
+	bne	$2, 0f		# U :
+	ret			# L0 :
+
+0:	or	$5, $18, $2	# E :
+	nop
+	bne	$2, 2f		# U :
+	and	$24, 0x80, $3	# E : no zero next byte
+
+	nop			# E :
+	bne	$3, 1f		# U :
+	/* Here there are bytes left in the current word.  Clear one.  */
+	addq	$24, $24, $24	# E : end-of-count bit <<= 1
+	nop			# E :
+
+2:	zap	$1, $24, $1	# U :
+	nop			# E :
+	stq_u	$1, 0($16)	# L :
+	ret			# L0 :
+
+1:	/* Here we must clear the first byte of the next DST word */
+	stb	$31, 8($16)	# L :
+	nop			# E :
+	nop			# E :
+	ret			# L0 :
+
+$zerocount:
+	nop			# E :
+	nop			# E :
+	nop			# E :
+	ret			# L0 :
+
+	.end strncat

arch/alpha/lib/ev67-strrchr.S

@@ -0,0 +1,109 @@
+/*
+ * arch/alpha/lib/ev67-strrchr.S
+ * 21264 version by Rick Gorton <rick.gorton@alpha-processor.com>
+ *
+ * Finds length of a 0-terminated string.  Optimized for the
+ * Alpha architecture:
+ *
+ *	- memory accessed as aligned quadwords only
+ *	- uses bcmpge to compare 8 bytes in parallel
+ *
+ * Much of the information about 21264 scheduling/coding comes from:
+ *	Compiler Writer's Guide for the Alpha 21264
+ *	abbreviated as 'CWG' in other comments here
+ *	ftp.digital.com/pub/Digital/info/semiconductor/literature/dsc-library.html
+ * Scheduling notation:
+ *	E	- either cluster
+ *	U	- upper subcluster; U0 - subcluster U0; U1 - subcluster U1
+ *	L	- lower subcluster; L0 - subcluster L0; L1 - subcluster L1
+ */
+
+
+#include <asm/regdef.h>
+
+	.set noreorder
+	.set noat
+
+	.align 4
+	.ent strrchr
+	.globl strrchr
+strrchr:
+	.frame sp, 0, ra
+	.prologue 0
+
+	and	a1, 0xff, t2	# E : 00000000000000ch
+	insbl	a1, 1, t4	# U : 000000000000ch00
+	insbl	a1, 2, t5	# U : 0000000000ch0000
+	ldq_u   t0, 0(a0)	# L : load first quadword Latency=3
+
+	mov	zero, t6	# E : t6 is last match aligned addr
+	or	t2, t4, a1	# E : 000000000000chch
+	sll	t5, 8, t3	# U : 00000000ch000000
+	mov	zero, t8	# E : t8 is last match byte compare mask
+
+	andnot  a0, 7, v0	# E : align source addr
+	or	t5, t3, t3	# E : 00000000chch0000
+	sll	a1, 32, t2	# U : 0000chch00000000
+	sll	a1, 48, t4	# U : chch000000000000
+
+	or	t4, a1, a1	# E : chch00000000chch
+	or	t2, t3, t2	# E : 0000chchchch0000
+	or	a1, t2, a1	# E : chchchchchchchch
+	lda	t5, -1		# E : build garbage mask
+
+	cmpbge  zero, t0, t1	# E : bits set iff byte == zero
+	mskqh	t5, a0, t4	# E : Complete garbage mask
+	xor	t0, a1, t2	# E : make bytes == c zero
+	cmpbge	zero, t4, t4	# E : bits set iff byte is garbage
+
+	cmpbge  zero, t2, t3	# E : bits set iff byte == c
+	andnot	t1, t4, t1	# E : clear garbage from null test
+	andnot	t3, t4, t3	# E : clear garbage from char test
+	bne	t1, $eos	# U : did we already hit the terminator?
+
+	/* Character search main loop */
+$loop:
+	ldq	t0, 8(v0)	# L : load next quadword
+	cmovne	t3, v0, t6	# E : save previous comparisons match
+	nop			#   : Latency=2, extra map slot (keep nop with cmov)
+	nop
+
+	cmovne	t3, t3, t8	# E : Latency=2, extra map slot
+	nop			#   : keep with cmovne
+	addq	v0, 8, v0	# E :
+	xor	t0, a1, t2	# E :
+
+	cmpbge	zero, t0, t1	# E : bits set iff byte == zero
+	cmpbge	zero, t2, t3	# E : bits set iff byte == c
+	beq	t1, $loop	# U : if we havnt seen a null, loop
+	nop
+
+	/* Mask out character matches after terminator */
+$eos:
+	negq	t1, t4		# E : isolate first null byte match
+	and	t1, t4, t4	# E :
+	subq	t4, 1, t5	# E : build a mask of the bytes up to...
+	or	t4, t5, t4	# E : ... and including the null
+
+	and	t3, t4, t3	# E : mask out char matches after null
+	cmovne	t3, t3, t8	# E : save it, if match found Latency=2, extra map slot
+	nop			#   : Keep with cmovne
+	nop
+
+	cmovne	t3, v0, t6	# E :
+	nop			#   : Keep with cmovne
+	/* Locate the address of the last matched character */
+	ctlz	t8, t2		# U0 : Latency=3 (0x40 for t8=0)
+	nop
+
+	cmoveq	t8, 0x3f, t2	# E : Compensate for case when no match is seen
+	nop			# E : hide the cmov latency (2) behind ctlz latency
+	lda	t5, 0x3f($31)	# E :
+	subq	t5, t2, t5	# E : Normalize leading zero count
+
+	addq	t6, t5, v0	# E : and add to quadword address
+	ret			# L0 : Latency=3
+	nop
+	nop
+
+	.end strrchr

arch/alpha/lib/fls.c

@@ -0,0 +1,38 @@
+/* 
+ * arch/alpha/lib/fls.c
+ */
+
+#include <linux/module.h>
+#include <linux/bitops.h>
+
+/* This is fls(x)-1, except zero is held to zero.  This allows most
+   efficient input into extbl, plus it allows easy handling of fls(0)=0.  */
+
+const unsigned char __flsm1_tab[256] = 
+{
+  0,
+  0,
+  1, 1,
+  2, 2, 2, 2,
+  3, 3, 3, 3, 3, 3, 3, 3,
+  4, 4, 4, 4, 4, 4, 4, 4, 4, 4, 4, 4, 4, 4, 4, 4,
+
+  5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5,
+  5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5,
+
+  6, 6, 6, 6, 6, 6, 6, 6, 6, 6, 6, 6, 6, 6, 6, 6,
+  6, 6, 6, 6, 6, 6, 6, 6, 6, 6, 6, 6, 6, 6, 6, 6,
+  6, 6, 6, 6, 6, 6, 6, 6, 6, 6, 6, 6, 6, 6, 6, 6,
+  6, 6, 6, 6, 6, 6, 6, 6, 6, 6, 6, 6, 6, 6, 6, 6,
+
+  7, 7, 7, 7, 7, 7, 7, 7, 7, 7, 7, 7, 7, 7, 7, 7,
+  7, 7, 7, 7, 7, 7, 7, 7, 7, 7, 7, 7, 7, 7, 7, 7,
+  7, 7, 7, 7, 7, 7, 7, 7, 7, 7, 7, 7, 7, 7, 7, 7,
+  7, 7, 7, 7, 7, 7, 7, 7, 7, 7, 7, 7, 7, 7, 7, 7,
+  7, 7, 7, 7, 7, 7, 7, 7, 7, 7, 7, 7, 7, 7, 7, 7,
+  7, 7, 7, 7, 7, 7, 7, 7, 7, 7, 7, 7, 7, 7, 7, 7,
+  7, 7, 7, 7, 7, 7, 7, 7, 7, 7, 7, 7, 7, 7, 7, 7,
+  7, 7, 7, 7, 7, 7, 7, 7, 7, 7, 7, 7, 7, 7, 7, 7,
+};
+
+EXPORT_SYMBOL(__flsm1_tab);

arch/alpha/lib/fpreg.c

@@ -0,0 +1,193 @@
+/*
+ * arch/alpha/lib/fpreg.c
+ *
+ * (C) Copyright 1998 Linus Torvalds
+ */
+
+#if defined(CONFIG_ALPHA_EV6) || defined(CONFIG_ALPHA_EV67)
+#define STT(reg,val)  asm volatile ("ftoit $f"#reg",%0" : "=r"(val));
+#else
+#define STT(reg,val)  asm volatile ("stt $f"#reg",%0" : "=m"(val));
+#endif
+
+unsigned long
+alpha_read_fp_reg (unsigned long reg)
+{
+	unsigned long val;
+
+	switch (reg) {
+	      case  0: STT( 0, val); break;
+	      case  1: STT( 1, val); break;
+	      case  2: STT( 2, val); break;
+	      case  3: STT( 3, val); break;
+	      case  4: STT( 4, val); break;
+	      case  5: STT( 5, val); break;
+	      case  6: STT( 6, val); break;
+	      case  7: STT( 7, val); break;
+	      case  8: STT( 8, val); break;
+	      case  9: STT( 9, val); break;
+	      case 10: STT(10, val); break;
+	      case 11: STT(11, val); break;
+	      case 12: STT(12, val); break;
+	      case 13: STT(13, val); break;
+	      case 14: STT(14, val); break;
+	      case 15: STT(15, val); break;
+	      case 16: STT(16, val); break;
+	      case 17: STT(17, val); break;
+	      case 18: STT(18, val); break;
+	      case 19: STT(19, val); break;
+	      case 20: STT(20, val); break;
+	      case 21: STT(21, val); break;
+	      case 22: STT(22, val); break;
+	      case 23: STT(23, val); break;
+	      case 24: STT(24, val); break;
+	      case 25: STT(25, val); break;
+	      case 26: STT(26, val); break;
+	      case 27: STT(27, val); break;
+	      case 28: STT(28, val); break;
+	      case 29: STT(29, val); break;
+	      case 30: STT(30, val); break;
+	      case 31: STT(31, val); break;
+	      default: return 0;
+	}
+	return val;
+}
+
+#if defined(CONFIG_ALPHA_EV6) || defined(CONFIG_ALPHA_EV67)
+#define LDT(reg,val)  asm volatile ("itoft %0,$f"#reg : : "r"(val));
+#else
+#define LDT(reg,val)  asm volatile ("ldt $f"#reg",%0" : : "m"(val));
+#endif
+
+void
+alpha_write_fp_reg (unsigned long reg, unsigned long val)
+{
+	switch (reg) {
+	      case  0: LDT( 0, val); break;
+	      case  1: LDT( 1, val); break;
+	      case  2: LDT( 2, val); break;
+	      case  3: LDT( 3, val); break;
+	      case  4: LDT( 4, val); break;
+	      case  5: LDT( 5, val); break;
+	      case  6: LDT( 6, val); break;
+	      case  7: LDT( 7, val); break;
+	      case  8: LDT( 8, val); break;
+	      case  9: LDT( 9, val); break;
+	      case 10: LDT(10, val); break;
+	      case 11: LDT(11, val); break;
+	      case 12: LDT(12, val); break;
+	      case 13: LDT(13, val); break;
+	      case 14: LDT(14, val); break;
+	      case 15: LDT(15, val); break;
+	      case 16: LDT(16, val); break;
+	      case 17: LDT(17, val); break;
+	      case 18: LDT(18, val); break;
+	      case 19: LDT(19, val); break;
+	      case 20: LDT(20, val); break;
+	      case 21: LDT(21, val); break;
+	      case 22: LDT(22, val); break;
+	      case 23: LDT(23, val); break;
+	      case 24: LDT(24, val); break;
+	      case 25: LDT(25, val); break;
+	      case 26: LDT(26, val); break;
+	      case 27: LDT(27, val); break;
+	      case 28: LDT(28, val); break;
+	      case 29: LDT(29, val); break;
+	      case 30: LDT(30, val); break;
+	      case 31: LDT(31, val); break;
+	}
+}
+
+#if defined(CONFIG_ALPHA_EV6) || defined(CONFIG_ALPHA_EV67)
+#define STS(reg,val)  asm volatile ("ftois $f"#reg",%0" : "=r"(val));
+#else
+#define STS(reg,val)  asm volatile ("sts $f"#reg",%0" : "=m"(val));
+#endif
+
+unsigned long
+alpha_read_fp_reg_s (unsigned long reg)
+{
+	unsigned long val;
+
+	switch (reg) {
+	      case  0: STS( 0, val); break;
+	      case  1: STS( 1, val); break;
+	      case  2: STS( 2, val); break;
+	      case  3: STS( 3, val); break;
+	      case  4: STS( 4, val); break;
+	      case  5: STS( 5, val); break;
+	      case  6: STS( 6, val); break;
+	      case  7: STS( 7, val); break;
+	      case  8: STS( 8, val); break;
+	      case  9: STS( 9, val); break;
+	      case 10: STS(10, val); break;
+	      case 11: STS(11, val); break;
+	      case 12: STS(12, val); break;
+	      case 13: STS(13, val); break;
+	      case 14: STS(14, val); break;
+	      case 15: STS(15, val); break;
+	      case 16: STS(16, val); break;
+	      case 17: STS(17, val); break;
+	      case 18: STS(18, val); break;
+	      case 19: STS(19, val); break;
+	      case 20: STS(20, val); break;
+	      case 21: STS(21, val); break;
+	      case 22: STS(22, val); break;
+	      case 23: STS(23, val); break;
+	      case 24: STS(24, val); break;
+	      case 25: STS(25, val); break;
+	      case 26: STS(26, val); break;
+	      case 27: STS(27, val); break;
+	      case 28: STS(28, val); break;
+	      case 29: STS(29, val); break;
+	      case 30: STS(30, val); break;
+	      case 31: STS(31, val); break;
+	      default: return 0;
+	}
+	return val;
+}
+
+#if defined(CONFIG_ALPHA_EV6) || defined(CONFIG_ALPHA_EV67)
+#define LDS(reg,val)  asm volatile ("itofs %0,$f"#reg : : "r"(val));
+#else
+#define LDS(reg,val)  asm volatile ("lds $f"#reg",%0" : : "m"(val));
+#endif
+
+void
+alpha_write_fp_reg_s (unsigned long reg, unsigned long val)
+{
+	switch (reg) {
+	      case  0: LDS( 0, val); break;
+	      case  1: LDS( 1, val); break;
+	      case  2: LDS( 2, val); break;
+	      case  3: LDS( 3, val); break;
+	      case  4: LDS( 4, val); break;
+	      case  5: LDS( 5, val); break;
+	      case  6: LDS( 6, val); break;
+	      case  7: LDS( 7, val); break;
+	      case  8: LDS( 8, val); break;
+	      case  9: LDS( 9, val); break;
+	      case 10: LDS(10, val); break;
+	      case 11: LDS(11, val); break;
+	      case 12: LDS(12, val); break;
+	      case 13: LDS(13, val); break;
+	      case 14: LDS(14, val); break;
+	      case 15: LDS(15, val); break;
+	      case 16: LDS(16, val); break;
+	      case 17: LDS(17, val); break;
+	      case 18: LDS(18, val); break;
+	      case 19: LDS(19, val); break;
+	      case 20: LDS(20, val); break;
+	      case 21: LDS(21, val); break;
+	      case 22: LDS(22, val); break;
+	      case 23: LDS(23, val); break;
+	      case 24: LDS(24, val); break;
+	      case 25: LDS(25, val); break;
+	      case 26: LDS(26, val); break;
+	      case 27: LDS(27, val); break;
+	      case 28: LDS(28, val); break;
+	      case 29: LDS(29, val); break;
+	      case 30: LDS(30, val); break;
+	      case 31: LDS(31, val); break;
+	}
+}

arch/alpha/lib/memchr.S

@@ -0,0 +1,164 @@
+/* Copyright (C) 1996 Free Software Foundation, Inc.
+   This file is part of the GNU C Library.
+   Contributed by David Mosberger (davidm@cs.arizona.edu).
+
+   The GNU C Library is free software; you can redistribute it and/or
+   modify it under the terms of the GNU Library General Public License as
+   published by the Free Software Foundation; either version 2 of the
+   License, or (at your option) any later version.
+
+   The GNU C Library 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
+   Library General Public License for more details.
+
+   You should have received a copy of the GNU Library General Public
+   License along with the GNU C Library; see the file COPYING.LIB.  If not,
+   write to the Free Software Foundation, Inc., 59 Temple Place - Suite 330,
+   Boston, MA 02111-1307, USA.  */
+
+/* Finds characters in a memory area.  Optimized for the Alpha:
+
+      - memory accessed as aligned quadwords only
+      - uses cmpbge to compare 8 bytes in parallel
+      - does binary search to find 0 byte in last
+        quadword (HAKMEM needed 12 instructions to
+        do this instead of the 9 instructions that
+        binary search needs).
+
+For correctness consider that:
+
+      - only minimum number of quadwords may be accessed
+      - the third argument is an unsigned long
+*/
+
+        .set noreorder
+        .set noat
+
+	.globl memchr
+	.ent memchr
+memchr:
+	.frame $30,0,$26,0
+	.prologue 0
+
+	# Hack -- if someone passes in (size_t)-1, hoping to just
+	# search til the end of the address space, we will overflow
+	# below when we find the address of the last byte.  Given
+	# that we will never have a 56-bit address space, cropping
+	# the length is the easiest way to avoid trouble.
+	zap	$18, 0x80, $5	#-e0	:
+
+	beq	$18, $not_found	# .. e1 :
+        ldq_u   $1, 0($16)	# e1	: load first quadword
+	insbl	$17, 1, $2	# .. e0 : $2 = 000000000000ch00
+	and	$17, 0xff, $17	#-e0    : $17 = 00000000000000ch
+	cmpult	$18, 9, $4	# .. e1 :
+	or	$2, $17, $17	# e0    : $17 = 000000000000chch
+        lda     $3, -1($31)	# .. e1 :
+	sll	$17, 16, $2	#-e0    : $2 = 00000000chch0000
+	addq	$16, $5, $5	# .. e1 :
+	or	$2, $17, $17	# e1    : $17 = 00000000chchchch
+	unop			#	:
+	sll	$17, 32, $2	#-e0    : $2 = chchchch00000000
+	or	$2, $17, $17	# e1	: $17 = chchchchchchchch
+	extql	$1, $16, $7	# e0    : 
+	beq	$4, $first_quad	# .. e1 :
+
+	ldq_u	$6, -1($5)	#-e1	: eight or less bytes to search
+	extqh	$6, $16, $6	# .. e0 :
+	mov	$16, $0		# e0	:
+	or	$7, $6, $1	# .. e1 : $1 = quadword starting at $16
+
+	# Deal with the case where at most 8 bytes remain to be searched
+	# in $1.  E.g.:
+	#	$18 = 6
+	#	$1 = ????c6c5c4c3c2c1
+$last_quad:
+	negq	$18, $6		#-e0	:
+        xor	$17, $1, $1	# .. e1 :
+	srl	$3, $6, $6	# e0    : $6 = mask of $18 bits set
+        cmpbge  $31, $1, $2	# .. e1 :
+	and	$2, $6, $2	#-e0	:
+        beq     $2, $not_found	# .. e1 :
+
+$found_it:
+	# Now, determine which byte matched:
+        negq    $2, $3		# e0	:
+        and     $2, $3, $2	# e1	:
+
+        and     $2, 0x0f, $1	#-e0	:
+        addq    $0, 4, $3	# .. e1 :
+        cmoveq  $1, $3, $0	# e0	:
+
+        addq    $0, 2, $3	# .. e1 :
+        and     $2, 0x33, $1	#-e0	:
+        cmoveq  $1, $3, $0	# .. e1 :
+
+        and     $2, 0x55, $1	# e0	:
+        addq    $0, 1, $3	# .. e1 :
+        cmoveq  $1, $3, $0	#-e0	:
+
+$done:	ret			# .. e1 :
+
+	# Deal with the case where $18 > 8 bytes remain to be
+	# searched.  $16 may not be aligned.
+	.align 4
+$first_quad:
+	andnot	$16, 0x7, $0	#-e1	:
+        insqh   $3, $16, $2	# .. e0	: $2 = 0000ffffffffffff ($16<0:2> ff)
+        xor	$1, $17, $1	# e0	:
+	or	$1, $2, $1	# e1	: $1 = ====ffffffffffff
+        cmpbge  $31, $1, $2	#-e0	:
+        bne     $2, $found_it	# .. e1 :
+
+	# At least one byte left to process.
+
+	ldq	$1, 8($0)	# e0	:
+	subq	$5, 1, $18	# .. e1 :
+	addq	$0, 8, $0	#-e0	:
+
+	# Make $18 point to last quad to be accessed (the
+	# last quad may or may not be partial).
+
+	andnot	$18, 0x7, $18	# .. e1 :
+	cmpult	$0, $18, $2	# e0	:
+	beq	$2, $final	# .. e1 :
+
+	# At least two quads remain to be accessed.
+
+	subq	$18, $0, $4	#-e0	: $4 <- nr quads to be processed
+	and	$4, 8, $4	# e1	: odd number of quads?
+	bne	$4, $odd_quad_count # e1 :
+
+	# At least three quads remain to be accessed
+
+	mov	$1, $4		# e0	: move prefetched value to correct reg
+
+	.align	4
+$unrolled_loop:
+	ldq	$1, 8($0)	#-e0	: prefetch $1
+	xor	$17, $4, $2	# .. e1 :
+	cmpbge	$31, $2, $2	# e0	:
+	bne	$2, $found_it	# .. e1 :
+
+	addq	$0, 8, $0	#-e0	:
+$odd_quad_count:
+	xor	$17, $1, $2	# .. e1 :
+	ldq	$4, 8($0)	# e0	: prefetch $4
+	cmpbge	$31, $2, $2	# .. e1 :
+	addq	$0, 8, $6	#-e0	:
+	bne	$2, $found_it	# .. e1	:
+
+	cmpult	$6, $18, $6	# e0	:
+	addq	$0, 8, $0	# .. e1 :
+	bne	$6, $unrolled_loop #-e1 :
+
+	mov	$4, $1		# e0	: move prefetched value into $1
+$final:	subq	$5, $0, $18	# .. e1	: $18 <- number of bytes left to do
+	bne	$18, $last_quad	# e1	:
+
+$not_found:
+	mov	$31, $0		#-e0	:
+	ret			# .. e1 :
+
+        .end memchr

arch/alpha/lib/memcpy.c

@@ -0,0 +1,163 @@
+/*
+ *  linux/arch/alpha/lib/memcpy.c
+ *
+ *  Copyright (C) 1995  Linus Torvalds
+ */
+
+/*
+ * This is a reasonably optimized memcpy() routine.
+ */
+
+/*
+ * Note that the C code is written to be optimized into good assembly. However,
+ * at this point gcc is unable to sanely compile "if (n >= 0)", resulting in a
+ * explicit compare against 0 (instead of just using the proper "blt reg, xx" or
+ * "bge reg, xx"). I hope alpha-gcc will be fixed to notice this eventually..
+ */
+
+#include <linux/types.h>
+
+/*
+ * This should be done in one go with ldq_u*2/mask/stq_u. Do it
+ * with a macro so that we can fix it up later..
+ */
+#define ALIGN_DEST_TO8_UP(d,s,n) \
+	while (d & 7) { \
+		if (n <= 0) return; \
+		n--; \
+		*(char *) d = *(char *) s; \
+		d++; s++; \
+	}
+#define ALIGN_DEST_TO8_DN(d,s,n) \
+	while (d & 7) { \
+		if (n <= 0) return; \
+		n--; \
+		d--; s--; \
+		*(char *) d = *(char *) s; \
+	}
+
+/*
+ * This should similarly be done with ldq_u*2/mask/stq. The destination
+ * is aligned, but we don't fill in a full quad-word
+ */
+#define DO_REST_UP(d,s,n) \
+	while (n > 0) { \
+		n--; \
+		*(char *) d = *(char *) s; \
+		d++; s++; \
+	}
+#define DO_REST_DN(d,s,n) \
+	while (n > 0) { \
+		n--; \
+		d--; s--; \
+		*(char *) d = *(char *) s; \
+	}
+
+/*
+ * This should be done with ldq/mask/stq. The source and destination are
+ * aligned, but we don't fill in a full quad-word
+ */
+#define DO_REST_ALIGNED_UP(d,s,n) DO_REST_UP(d,s,n)
+#define DO_REST_ALIGNED_DN(d,s,n) DO_REST_DN(d,s,n)
+
+/*
+ * This does unaligned memory copies. We want to avoid storing to
+ * an unaligned address, as that would do a read-modify-write cycle.
+ * We also want to avoid double-reading the unaligned reads.
+ *
+ * Note the ordering to try to avoid load (and address generation) latencies.
+ */
+static inline void __memcpy_unaligned_up (unsigned long d, unsigned long s,
+					  long n)
+{
+	ALIGN_DEST_TO8_UP(d,s,n);
+	n -= 8;			/* to avoid compare against 8 in the loop */
+	if (n >= 0) {
+		unsigned long low_word, high_word;
+		__asm__("ldq_u %0,%1":"=r" (low_word):"m" (*(unsigned long *) s));
+		do {
+			unsigned long tmp;
+			__asm__("ldq_u %0,%1":"=r" (high_word):"m" (*(unsigned long *)(s+8)));
+			n -= 8;
+			__asm__("extql %1,%2,%0"
+				:"=r" (low_word)
+				:"r" (low_word), "r" (s));
+			__asm__("extqh %1,%2,%0"
+				:"=r" (tmp)
+				:"r" (high_word), "r" (s));
+			s += 8;
+			*(unsigned long *) d = low_word | tmp;
+			d += 8;
+			low_word = high_word;
+		} while (n >= 0);
+	}
+	n += 8;
+	DO_REST_UP(d,s,n);
+}
+
+static inline void __memcpy_unaligned_dn (unsigned long d, unsigned long s,
+					  long n)
+{
+	/* I don't understand AXP assembler well enough for this. -Tim */
+	s += n;
+	d += n;
+	while (n--)
+		* (char *) --d = * (char *) --s;
+}
+
+/*
+ * Hmm.. Strange. The __asm__ here is there to make gcc use an integer register
+ * for the load-store. I don't know why, but it would seem that using a floating
+ * point register for the move seems to slow things down (very small difference,
+ * though).
+ *
+ * Note the ordering to try to avoid load (and address generation) latencies.
+ */
+static inline void __memcpy_aligned_up (unsigned long d, unsigned long s,
+					long n)
+{
+	ALIGN_DEST_TO8_UP(d,s,n);
+	n -= 8;
+	while (n >= 0) {
+		unsigned long tmp;
+		__asm__("ldq %0,%1":"=r" (tmp):"m" (*(unsigned long *) s));
+		n -= 8;
+		s += 8;
+		*(unsigned long *) d = tmp;
+		d += 8;
+	}
+	n += 8;
+	DO_REST_ALIGNED_UP(d,s,n);
+}
+static inline void __memcpy_aligned_dn (unsigned long d, unsigned long s,
+					long n)
+{
+	s += n;
+	d += n;
+	ALIGN_DEST_TO8_DN(d,s,n);
+	n -= 8;
+	while (n >= 0) {
+		unsigned long tmp;
+		s -= 8;
+		__asm__("ldq %0,%1":"=r" (tmp):"m" (*(unsigned long *) s));
+		n -= 8;
+		d -= 8;
+		*(unsigned long *) d = tmp;
+	}
+	n += 8;
+	DO_REST_ALIGNED_DN(d,s,n);
+}
+
+void * memcpy(void * dest, const void *src, size_t n)
+{
+	if (!(((unsigned long) dest ^ (unsigned long) src) & 7)) {
+		__memcpy_aligned_up ((unsigned long) dest, (unsigned long) src,
+				     n);
+		return dest;
+	}
+	__memcpy_unaligned_up ((unsigned long) dest, (unsigned long) src, n);
+	return dest;
+}
+
+/* For backward modules compatibility, define __memcpy.  */
+asm("__memcpy = memcpy; .globl __memcpy");

arch/alpha/lib/memmove.S

@@ -0,0 +1,181 @@
+/*
+ * arch/alpha/lib/memmove.S
+ *
+ * Barely optimized memmove routine for Alpha EV5.
+ *
+ * This is hand-massaged output from the original memcpy.c.  We defer to
+ * memcpy whenever possible; the backwards copy loops are not unrolled.
+ */
+        
+	.set noat
+	.set noreorder
+	.text
+
+	.align 4
+	.globl memmove
+	.ent memmove
+memmove:
+	ldgp $29, 0($27)
+	unop
+	nop
+	.prologue 1
+
+	addq $16,$18,$4
+	addq $17,$18,$5
+	cmpule $4,$17,$1		/*  dest + n <= src  */
+	cmpule $5,$16,$2		/*  dest >= src + n  */
+
+	bis $1,$2,$1
+	mov $16,$0
+	xor $16,$17,$2
+	bne $1,memcpy			!samegp
+
+	and $2,7,$2			/* Test for src/dest co-alignment.  */
+	and $16,7,$1
+	cmpule $16,$17,$3
+	bne $3,$memmove_up		/* dest < src */
+
+	and $4,7,$1
+	bne $2,$misaligned_dn
+	unop
+	beq $1,$skip_aligned_byte_loop_head_dn
+
+$aligned_byte_loop_head_dn:
+	lda $4,-1($4)
+	lda $5,-1($5)
+	unop
+	ble $18,$egress
+
+	ldq_u $3,0($5)
+	ldq_u $2,0($4)
+	lda $18,-1($18)
+	extbl $3,$5,$1
+
+	insbl $1,$4,$1
+	mskbl $2,$4,$2
+	bis $1,$2,$1
+	and $4,7,$6
+
+	stq_u $1,0($4)
+	bne $6,$aligned_byte_loop_head_dn
+
+$skip_aligned_byte_loop_head_dn:
+	lda $18,-8($18)
+	blt $18,$skip_aligned_word_loop_dn
+
+$aligned_word_loop_dn:
+	ldq $1,-8($5)
+	nop
+	lda $5,-8($5)
+	lda $18,-8($18)
+
+	stq $1,-8($4)
+	nop
+	lda $4,-8($4)
+	bge $18,$aligned_word_loop_dn
+
+$skip_aligned_word_loop_dn:
+	lda $18,8($18)
+	bgt $18,$byte_loop_tail_dn
+	unop
+	ret $31,($26),1
+
+	.align 4
+$misaligned_dn:
+	nop
+	fnop
+	unop
+	beq $18,$egress
+
+$byte_loop_tail_dn:
+	ldq_u $3,-1($5)
+	ldq_u $2,-1($4)
+	lda $5,-1($5)
+	lda $4,-1($4)
+
+	lda $18,-1($18)
+	extbl $3,$5,$1
+	insbl $1,$4,$1
+	mskbl $2,$4,$2
+
+	bis $1,$2,$1
+	stq_u $1,0($4)
+	bgt $18,$byte_loop_tail_dn
+	br $egress
+
+$memmove_up:
+	mov $16,$4
+	mov $17,$5
+	bne $2,$misaligned_up
+	beq $1,$skip_aligned_byte_loop_head_up
+
+$aligned_byte_loop_head_up:
+	unop
+	ble $18,$egress
+	ldq_u $3,0($5)
+	ldq_u $2,0($4)
+
+	lda $18,-1($18)
+	extbl $3,$5,$1
+	insbl $1,$4,$1
+	mskbl $2,$4,$2
+
+	bis $1,$2,$1
+	lda $5,1($5)
+	stq_u $1,0($4)
+	lda $4,1($4)
+
+	and $4,7,$6
+	bne $6,$aligned_byte_loop_head_up
+
+$skip_aligned_byte_loop_head_up:
+	lda $18,-8($18)
+	blt $18,$skip_aligned_word_loop_up
+
+$aligned_word_loop_up:
+	ldq $1,0($5)
+	nop
+	lda $5,8($5)
+	lda $18,-8($18)
+
+	stq $1,0($4)
+	nop
+	lda $4,8($4)
+	bge $18,$aligned_word_loop_up
+
+$skip_aligned_word_loop_up:
+	lda $18,8($18)
+	bgt $18,$byte_loop_tail_up
+	unop
+	ret $31,($26),1
+
+	.align 4
+$misaligned_up:
+	nop
+	fnop
+	unop
+	beq $18,$egress
+
+$byte_loop_tail_up:
+	ldq_u $3,0($5)
+	ldq_u $2,0($4)
+	lda $18,-1($18)
+	extbl $3,$5,$1
+
+	insbl $1,$4,$1
+	mskbl $2,$4,$2
+	bis $1,$2,$1
+	stq_u $1,0($4)
+
+	lda $5,1($5)
+	lda $4,1($4)
+	nop
+	bgt $18,$byte_loop_tail_up
+
+$egress:
+	ret $31,($26),1
+	nop
+	nop
+	nop
+
+	.end memmove

arch/alpha/lib/memset.S

@@ -0,0 +1,127 @@
+/*
+ * linux/arch/alpha/lib/memset.S
+ *
+ * This is an efficient (and small) implementation of the C library "memset()"
+ * function for the alpha.
+ *
+ *	(C) Copyright 1996 Linus Torvalds
+ *
+ * This routine is "moral-ware": you are free to use it any way you wish, and
+ * the only obligation I put on you is a moral one: if you make any improvements
+ * to the routine, please send me your improvements for me to use similarly.
+ *
+ * The scheduling comments are according to the EV5 documentation (and done by 
+ * hand, so they might well be incorrect, please do tell me about it..)
+ */
+
+	.set noat
+	.set noreorder
+.text
+	.globl memset
+	.globl __memset
+	.globl ___memset
+	.globl __memsetw
+	.globl __constant_c_memset
+
+	.ent ___memset
+.align 5
+___memset:
+	.frame $30,0,$26,0
+	.prologue 0
+
+	and $17,255,$1		/* E1 */
+	insbl $17,1,$17		/* .. E0 */
+	bis $17,$1,$17		/* E0 (p-c latency, next cycle) */
+	sll $17,16,$1		/* E1 (p-c latency, next cycle) */
+
+	bis $17,$1,$17		/* E0 (p-c latency, next cycle) */
+	sll $17,32,$1		/* E1 (p-c latency, next cycle) */
+	bis $17,$1,$17		/* E0 (p-c latency, next cycle) */
+	ldq_u $31,0($30)	/* .. E1 */
+
+.align 5
+__constant_c_memset:
+	addq $18,$16,$6		/* E0 */
+	bis $16,$16,$0		/* .. E1 */
+	xor $16,$6,$1		/* E0 */
+	ble $18,end		/* .. E1 */
+
+	bic $1,7,$1		/* E0 */
+	beq $1,within_one_quad	/* .. E1 (note EV5 zero-latency forwarding) */
+	and $16,7,$3		/* E0 */
+	beq $3,aligned		/* .. E1 (note EV5 zero-latency forwarding) */
+
+	ldq_u $4,0($16)		/* E0 */
+	bis $16,$16,$5		/* .. E1 */
+	insql $17,$16,$2	/* E0 */
+	subq $3,8,$3		/* .. E1 */
+
+	addq $18,$3,$18		/* E0	$18 is new count ($3 is negative) */
+	mskql $4,$16,$4		/* .. E1 (and possible load stall) */
+	subq $16,$3,$16		/* E0 	$16 is new aligned destination */
+	bis $2,$4,$1		/* .. E1 */
+
+	bis $31,$31,$31		/* E0 */
+	ldq_u $31,0($30)	/* .. E1 */
+	stq_u $1,0($5)		/* E0 */
+	bis $31,$31,$31		/* .. E1 */
+
+.align 4
+aligned:
+	sra $18,3,$3		/* E0 */
+	and $18,7,$18		/* .. E1 */
+	bis $16,$16,$5		/* E0 */
+	beq $3,no_quad		/* .. E1 */
+
+.align 3
+loop:
+	stq $17,0($5)		/* E0 */
+	subq $3,1,$3		/* .. E1 */
+	addq $5,8,$5		/* E0 */
+	bne $3,loop		/* .. E1 */
+
+no_quad:
+	bis $31,$31,$31		/* E0 */
+	beq $18,end		/* .. E1 */
+	ldq $7,0($5)		/* E0 */
+	mskqh $7,$6,$2		/* .. E1 (and load stall) */
+
+	insqh $17,$6,$4		/* E0 */
+	bis $2,$4,$1		/* .. E1 */
+	stq $1,0($5)		/* E0 */
+	ret $31,($26),1		/* .. E1 */
+
+.align 3
+within_one_quad:
+	ldq_u $1,0($16)		/* E0 */
+	insql $17,$16,$2	/* E1 */
+	mskql $1,$16,$4		/* E0 (after load stall) */
+	bis $2,$4,$2		/* E0 */
+
+	mskql $2,$6,$4		/* E0 */
+	mskqh $1,$6,$2		/* .. E1 */
+	bis $2,$4,$1		/* E0 */
+	stq_u $1,0($16)		/* E0 */
+
+end:
+	ret $31,($26),1		/* E1 */
+	.end ___memset
+
+	.align 5
+	.ent __memsetw
+__memsetw:
+	.prologue 0
+
+	inswl $17,0,$1		/* E0 */
+	inswl $17,2,$2		/* E0 */
+	inswl $17,4,$3		/* E0 */
+	or $1,$2,$1		/* .. E1 */
+	inswl $17,6,$4		/* E0 */
+	or $1,$3,$1		/* .. E1 */
+	or $1,$4,$17		/* E0 */
+	br __constant_c_memset	/* .. E1 */
+
+	.end __memsetw
+
+memset = ___memset
+__memset = ___memset

arch/alpha/lib/srm_printk.c

@@ -0,0 +1,41 @@
+/*
+ *	arch/alpha/lib/srm_printk.c
+ */
+
+#include <linux/kernel.h>
+#include <asm/console.h>
+
+long
+srm_printk(const char *fmt, ...)
+{
+	static char buf[1024];
+	va_list args;
+	long len, num_lf;
+	char *src, *dst;
+
+	va_start(args, fmt);
+	len = vsprintf(buf, fmt, args);
+	va_end(args);
+
+	/* count number of linefeeds in string: */
+
+	num_lf = 0;
+	for (src = buf; *src; ++src) {
+		if (*src == '\n') {
+			++num_lf;
+		}
+	}
+
+	if (num_lf) {
+		/* expand each linefeed into carriage-return/linefeed: */
+		for (dst = src + num_lf; src >= buf; ) {
+			if (*src == '\n') {
+				*dst-- = '\r';
+			}
+			*dst-- = *src--;
+		}
+	}
+
+	srm_puts(buf, num_lf+len);	
+        return len;
+}

arch/alpha/lib/srm_puts.c

@@ -0,0 +1,23 @@
+/*
+ *	arch/alpha/lib/srm_puts.c
+ */
+
+#include <linux/string.h>
+#include <asm/console.h>
+
+long
+srm_puts(const char *str, long len)
+{
+	long remaining, written;
+
+	if (!callback_init_done)
+		return len;
+
+	for (remaining = len; remaining > 0; remaining -= written)
+	{
+		written = callback_puts(0, str, remaining);
+		written &= 0xffffffff;
+		str += written;
+	}
+	return len;
+}

arch/alpha/lib/stacktrace.c

@@ -0,0 +1,102 @@
+#include <linux/kernel.h>
+
+typedef unsigned int instr;
+
+#define MAJOR_OP	0xfc000000
+#define LDA_OP		0x20000000
+#define STQ_OP		0xb4000000
+#define BR_OP		0xc0000000
+
+#define STK_ALLOC_1	0x23de8000 /* lda $30,-X($30) */
+#define STK_ALLOC_1M	0xffff8000
+#define STK_ALLOC_2	0x43c0153e /* subq $30,X,$30 */
+#define STK_ALLOC_2M	0xffe01fff
+
+#define MEM_REG		0x03e00000
+#define MEM_BASE	0x001f0000
+#define MEM_OFF		0x0000ffff
+#define MEM_OFF_SIGN	0x00008000
+#define	BASE_SP		0x001e0000
+
+#define STK_ALLOC_MATCH(INSTR)			\
+  (((INSTR) & STK_ALLOC_1M) == STK_ALLOC_1	\
+   || ((INSTR) & STK_ALLOC_2M) == STK_ALLOC_2)
+#define STK_PUSH_MATCH(INSTR) \
+  (((INSTR) & (MAJOR_OP | MEM_BASE | MEM_OFF_SIGN)) == (STQ_OP | BASE_SP))
+#define MEM_OP_OFFSET(INSTR) \
+  (((long)((INSTR) & MEM_OFF) << 48) >> 48)
+#define MEM_OP_REG(INSTR) \
+  (((INSTR) & MEM_REG) >> 22)
+
+/* Branches, jumps, PAL calls, and illegal opcodes end a basic block. */
+#define BB_END(INSTR)						\
+  (((instr)(INSTR) >= BR_OP) | ((instr)(INSTR) < LDA_OP) |	\
+   ((((instr)(INSTR) ^ 0x60000000) < 0x20000000) &		\
+    (((instr)(INSTR) & 0x0c000000) != 0)))
+
+#define IS_KERNEL_TEXT(PC) ((unsigned long)(PC) > START_ADDR)
+
+static char reg_name[][4] = {
+	"v0 ", "t0 ", "t1 ", "t2 ", "t3 ", "t4 ", "t5 ", "t6 ", "t7 ",
+	"s0 ", "s1 ", "s2 ", "s3 ", "s4 ", "s5 ", "s6 ", "a0 ", "a1 ",
+	"a2 ", "a3 ", "a4 ", "a5 ", "t8 ", "t9 ", "t10", "t11", "ra ",
+	"pv ", "at ", "gp ", "sp ", "0"
+};
+
+
+static instr *
+display_stored_regs(instr * pro_pc, unsigned char * sp)
+{
+	instr * ret_pc = 0;
+	int reg;
+	unsigned long value;
+
+	printk("Prologue [<%p>], Frame %p:\n", pro_pc, sp);
+	while (!BB_END(*pro_pc))
+		if (STK_PUSH_MATCH(*pro_pc)) {
+			reg = (*pro_pc & MEM_REG) >> 21;
+			value = *(unsigned long *)(sp + (*pro_pc & MEM_OFF));
+			if (reg == 26)
+				ret_pc = (instr *)value;
+			printk("\t\t%s / 0x%016lx\n", reg_name[reg], value);
+		}
+	return ret_pc;
+}
+
+static instr *
+seek_prologue(instr * pc)
+{
+	while (!STK_ALLOC_MATCH(*pc))
+		--pc;
+	while (!BB_END(*(pc - 1)))
+		--pc;
+	return pc;
+}
+
+static long
+stack_increment(instr * prologue_pc)
+{
+	while (!STK_ALLOC_MATCH(*prologue_pc))
+		++prologue_pc;
+
+	/* Count the bytes allocated. */
+	if ((*prologue_pc & STK_ALLOC_1M) == STK_ALLOC_1M)
+		return -(((long)(*prologue_pc) << 48) >> 48);
+	else
+		return (*prologue_pc >> 13) & 0xff;
+}
+
+void
+stacktrace(void)
+{
+	instr * ret_pc;
+	instr * prologue = (instr *)stacktrace;
+	register unsigned char * sp __asm__ ("$30");
+
+	printk("\tstack trace:\n");
+	do {
+		ret_pc = display_stored_regs(prologue, sp);
+		sp += stack_increment(prologue);
+		prologue = seek_prologue(ret_pc);
+	} while (IS_KERNEL_TEXT(ret_pc));
+}

arch/alpha/lib/strcat.S

@@ -0,0 +1,52 @@
+/*
+ * arch/alpha/lib/strcat.S
+ * Contributed by Richard Henderson (rth@tamu.edu)
+ *
+ * Append a null-terminated string from SRC to DST.
+ */
+
+	.text
+
+	.align 3
+	.globl strcat
+	.ent strcat
+strcat:
+	.frame $30, 0, $26
+	.prologue 0
+
+	mov	$16, $0		# set up return value
+
+	/* Find the end of the string.  */
+
+	ldq_u   $1, 0($16)	# load first quadword (a0 may be misaligned)
+	lda     $2, -1
+	insqh   $2, $16, $2
+	andnot  $16, 7, $16
+	or      $2, $1, $1
+	cmpbge  $31, $1, $2	# bits set iff byte == 0
+	bne     $2, $found
+
+$loop:	ldq     $1, 8($16)
+	addq    $16, 8, $16
+	cmpbge  $31, $1, $2
+	beq     $2, $loop
+
+$found:	negq    $2, $3		# clear all but least set bit
+	and     $2, $3, $2
+
+	and     $2, 0xf0, $3	# binary search for that set bit
+	and	$2, 0xcc, $4
+	and	$2, 0xaa, $5
+	cmovne	$3, 4, $3
+	cmovne	$4, 2, $4
+	cmovne	$5, 1, $5
+	addq	$3, $4, $3
+	addq	$16, $5, $16
+	addq	$16, $3, $16
+
+	/* Now do the append.  */
+
+	mov	$26, $23
+	br	__stxcpy
+
+	.end strcat

arch/alpha/lib/strchr.S

@@ -0,0 +1,70 @@
+/*
+ * arch/alpha/lib/strchr.S
+ * Contributed by Richard Henderson (rth@tamu.edu)
+ *
+ * Return the address of a given character within a null-terminated
+ * string, or null if it is not found.
+ */
+
+#include <asm/regdef.h>
+
+	.set noreorder
+	.set noat
+
+	.align 3
+	.globl strchr
+	.ent strchr
+strchr:
+	.frame sp, 0, ra
+	.prologue 0
+
+	zapnot	a1, 1, a1	# e0    : zero extend the search character
+	ldq_u   t0, 0(a0)	# .. e1 : load first quadword
+	sll	a1, 8, t5	# e0    : replicate the search character
+	andnot  a0, 7, v0	# .. e1 : align our loop pointer
+	or	t5, a1, a1	# e0    :
+	lda	t4, -1		# .. e1 : build garbage mask
+	sll	a1, 16, t5	# e0    :
+	cmpbge  zero, t0, t2	# .. e1 : bits set iff byte == zero
+	mskqh	t4, a0, t4	# e0    :
+	or	t5, a1, a1	# .. e1 :
+	sll	a1, 32, t5	# e0    :
+	cmpbge	zero, t4, t4	# .. e1 : bits set iff byte is garbage
+	or	t5, a1, a1	# e0    :
+	xor	t0, a1, t1	# .. e1 : make bytes == c zero
+	cmpbge  zero, t1, t3	# e0    : bits set iff byte == c
+	or	t2, t3, t0	# e1    : bits set iff char match or zero match
+	andnot	t0, t4, t0	# e0    : clear garbage bits
+	bne	t0, $found	# .. e1 (zdb)
+
+$loop:	ldq	t0, 8(v0)	# e0    :
+	addq	v0, 8, v0	# .. e1 :
+	nop			# e0    :
+	xor	t0, a1, t1	# .. e1 (ev5 data stall)
+	cmpbge	zero, t0, t2	# e0    : bits set iff byte == 0
+	cmpbge	zero, t1, t3	# .. e1 : bits set iff byte == c
+	or	t2, t3, t0	# e0    :
+	beq	t0, $loop	# .. e1 (zdb)
+
+$found:	negq    t0, t1		# e0    : clear all but least set bit
+	and     t0, t1, t0	# e1 (stall)
+
+	and	t0, t3, t1	# e0    : bit set iff byte was the char
+	beq	t1, $retnull	# .. e1 (zdb)
+
+	and     t0, 0xf0, t2	# e0    : binary search for that set bit
+	and	t0, 0xcc, t3	# .. e1 :
+	and	t0, 0xaa, t4	# e0    :
+	cmovne	t2, 4, t2	# .. e1 :
+	cmovne	t3, 2, t3	# e0    :
+	cmovne	t4, 1, t4	# .. e1 :
+	addq	t2, t3, t2	# e0    :
+	addq	v0, t4, v0	# .. e1 :
+	addq	v0, t2, v0	# e0    :
+	ret			# .. e1 :
+
+$retnull:
+	mov	zero, v0	# e0    :
+	ret			# .. e1 :
+
+	.end strchr

arch/alpha/lib/strcpy.S

@@ -0,0 +1,23 @@
+/*
+ * arch/alpha/lib/strcpy.S
+ * Contributed by Richard Henderson (rth@tamu.edu)
+ *
+ * Copy a null-terminated string from SRC to DST.  Return a pointer
+ * to the null-terminator in the source.
+ */
+
+	.text
+
+	.align 3
+	.globl strcpy
+	.ent strcpy
+strcpy:
+	.frame $30, 0, $26
+	.prologue 0
+
+	mov	$16, $0		# set up return value
+	mov	$26, $23	# set up return address
+	unop
+	br	__stxcpy	# do the copy
+
+	.end strcpy

arch/alpha/lib/strlen.S

@@ -0,0 +1,57 @@
+/*
+ * strlen.S (c) 1995 David Mosberger (davidm@cs.arizona.edu)
+ *
+ * Finds length of a 0-terminated string.  Optimized for the
+ * Alpha architecture:
+ *
+ *	- memory accessed as aligned quadwords only
+ *	- uses bcmpge to compare 8 bytes in parallel
+ *	- does binary search to find 0 byte in last
+ *	  quadword (HAKMEM needed 12 instructions to
+ *	  do this instead of the 9 instructions that
+ *	  binary search needs).
+ */
+
+	.set noreorder
+	.set noat
+
+	.align 3
+
+	.globl	strlen
+	.ent	strlen
+
+strlen:
+	ldq_u	$1, 0($16)	# load first quadword ($16  may be misaligned)
+	lda	$2, -1($31)
+	insqh	$2, $16, $2
+	andnot	$16, 7, $0
+	or	$2, $1, $1
+	cmpbge	$31, $1, $2	# $2  <- bitmask: bit i == 1 <==> i-th byte == 0
+	bne	$2, found
+
+loop:	ldq	$1, 8($0)
+	addq	$0, 8, $0	# addr += 8
+	nop			# helps dual issue last two insns
+	cmpbge	$31, $1, $2
+	beq	$2, loop
+
+found:	blbs	$2, done	# make aligned case fast
+	negq	$2, $3
+	and	$2, $3, $2
+
+	and	$2, 0x0f, $1
+	addq	$0, 4, $3
+	cmoveq	$1, $3, $0
+
+	and	$2, 0x33, $1
+	addq	$0, 2, $3
+	cmoveq	$1, $3, $0
+
+	and	$2, 0x55, $1
+	addq	$0, 1, $3
+	cmoveq	$1, $3, $0
+
+done:	subq	$0, $16, $0
+	ret	$31, ($26)
+
+	.end	strlen

arch/alpha/lib/strncat.S

@@ -0,0 +1,84 @@
+/*
+ * arch/alpha/lib/strncat.S
+ * Contributed by Richard Henderson (rth@tamu.edu)
+ *
+ * Append no more than COUNT characters from the null-terminated string SRC
+ * to the null-terminated string DST.  Always null-terminate the new DST.
+ *
+ * This differs slightly from the semantics in libc in that we never write
+ * past count, whereas libc may write to count+1.  This follows the generic
+ * implementation in lib/string.c and is, IMHO, more sensible.
+ */
+
+	.text
+
+	.align 3
+	.globl strncat
+	.ent strncat
+strncat:
+	.frame $30, 0, $26
+	.prologue 0
+
+	mov	$16, $0		# set up return value
+	beq	$18, $zerocount
+
+	/* Find the end of the string.  */
+
+	ldq_u   $1, 0($16)	# load first quadword ($16 may be misaligned)
+	lda     $2, -1($31)
+	insqh   $2, $16, $2
+	andnot  $16, 7, $16
+	or      $2, $1, $1
+	cmpbge  $31, $1, $2	# bits set iff byte == 0
+	bne     $2, $found
+
+$loop:	ldq     $1, 8($16)
+	addq    $16, 8, $16
+	cmpbge  $31, $1, $2
+	beq     $2, $loop
+
+$found:	negq    $2, $3		# clear all but least set bit
+	and     $2, $3, $2
+
+	and     $2, 0xf0, $3	# binary search for that set bit
+	and	$2, 0xcc, $4
+	and	$2, 0xaa, $5
+	cmovne	$3, 4, $3
+	cmovne	$4, 2, $4
+	cmovne	$5, 1, $5
+	addq	$3, $4, $3
+	addq	$16, $5, $16
+	addq	$16, $3, $16
+
+	/* Now do the append.  */
+
+	bsr	$23, __stxncpy
+
+	/* Worry about the null termination.  */
+
+	zapnot	$1, $27, $2	# was last byte a null?
+	bne	$2, 0f
+	ret
+
+0:	cmplt	$27, $24, $2	# did we fill the buffer completely?
+	or	$2, $18, $2
+	bne	$2, 2f
+
+	and	$24, 0x80, $2	# no zero next byte
+	bne	$2, 1f
+
+	/* Here there are bytes left in the current word.  Clear one.  */
+	addq	$24, $24, $24	# end-of-count bit <<= 1
+2:	zap	$1, $24, $1
+	stq_u	$1, 0($16)
+	ret
+
+1:	/* Here we must read the next DST word and clear the first byte.  */
+	ldq_u	$1, 8($16)
+	zap	$1, 1, $1
+	stq_u	$1, 8($16)
+
+$zerocount:
+	ret
+
+	.end strncat

arch/alpha/lib/strncpy.S

@@ -0,0 +1,81 @@
+/*
+ * arch/alpha/lib/strncpy.S
+ * Contributed by Richard Henderson (rth@tamu.edu)
+ *
+ * Copy no more than COUNT bytes of the null-terminated string from
+ * SRC to DST.  If SRC does not cover all of COUNT, the balance is
+ * zeroed.
+ *
+ * Or, rather, if the kernel cared about that weird ANSI quirk.  This
+ * version has cropped that bit o' nastiness as well as assuming that
+ * __stxncpy is in range of a branch.
+ */
+
+	.set noat
+	.set noreorder
+
+	.text
+
+	.align 4
+	.globl strncpy
+	.ent strncpy
+strncpy:
+	.frame $30, 0, $26
+	.prologue 0
+
+	mov	$16, $0		# set return value now
+	beq	$18, $zerolen
+	unop
+	bsr	$23, __stxncpy	# do the work of the copy
+
+	unop
+	bne	$18, $multiword	# do we have full words left?
+	subq	$24, 1, $3	# nope
+	subq	$27, 1, $4
+
+	or	$3, $24, $3	# clear the bits between the last
+	or	$4, $27, $4	# written byte and the last byte in COUNT
+	andnot	$3, $4, $4
+	zap	$1, $4, $1
+
+	stq_u	$1, 0($16)
+	ret
+
+	.align	4
+$multiword:
+	subq	$27, 1, $2	# clear the final bits in the prev word
+	or	$2, $27, $2
+	zapnot	$1, $2, $1
+	subq	$18, 1, $18
+
+	stq_u	$1, 0($16)
+	addq	$16, 8, $16
+	unop
+	beq	$18, 1f
+
+	nop
+	unop
+	nop
+	blbc	$18, 0f
+
+	stq_u	$31, 0($16)	# zero one word
+	subq	$18, 1, $18
+	addq	$16, 8, $16
+	beq	$18, 1f
+
+0:	stq_u	$31, 0($16)	# zero two words
+	subq	$18, 2, $18
+	stq_u	$31, 8($16)
+	addq	$16, 16, $16
+	bne	$18, 0b
+
+1:	ldq_u	$1, 0($16)	# clear the leading bits in the final word
+	subq	$24, 1, $2
+	or	$2, $24, $2
+
+	zap	$1, $2, $1
+	stq_u	$1, 0($16)
+$zerolen:
+	ret
+
+	.end	strncpy

arch/alpha/lib/strrchr.S

@@ -0,0 +1,87 @@
+/*
+ * arch/alpha/lib/strrchr.S
+ * Contributed by Richard Henderson (rth@tamu.edu)
+ *
+ * Return the address of the last occurrence of a given character
+ * within a null-terminated string, or null if it is not found.
+ */
+
+#include <asm/regdef.h>
+
+	.set noreorder
+	.set noat
+
+	.align 3
+	.ent strrchr
+	.globl strrchr
+strrchr:
+	.frame sp, 0, ra
+	.prologue 0
+
+	zapnot	a1, 1, a1	# e0    : zero extend our test character
+	mov	zero, t6	# .. e1 : t6 is last match aligned addr
+	sll	a1, 8, t5	# e0    : replicate our test character
+	mov	zero, t8	# .. e1 : t8 is last match byte compare mask
+	or	t5, a1, a1	# e0    :
+	ldq_u   t0, 0(a0)	# .. e1 : load first quadword
+	sll	a1, 16, t5	# e0    :
+	andnot  a0, 7, v0	# .. e1 : align source addr
+	or	t5, a1, a1	# e0    :
+	lda	t4, -1		# .. e1 : build garbage mask
+	sll	a1, 32, t5	# e0    :
+	cmpbge  zero, t0, t1	# .. e1 : bits set iff byte == zero
+	mskqh	t4, a0, t4	# e0    :
+	or	t5, a1, a1	# .. e1 : character replication complete
+	xor	t0, a1, t2	# e0    : make bytes == c zero
+	cmpbge	zero, t4, t4	# .. e1 : bits set iff byte is garbage
+	cmpbge  zero, t2, t3	# e0    : bits set iff byte == c
+	andnot	t1, t4, t1	# .. e1 : clear garbage from null test
+	andnot	t3, t4, t3	# e0    : clear garbage from char test
+	bne	t1, $eos	# .. e1 : did we already hit the terminator?
+
+	/* Character search main loop */
+$loop:
+	ldq	t0, 8(v0)	# e0    : load next quadword
+	cmovne	t3, v0, t6	# .. e1 : save previous comparisons match
+	cmovne	t3, t3, t8	# e0    :
+	addq	v0, 8, v0	# .. e1 :
+	xor	t0, a1, t2	# e0    :
+	cmpbge	zero, t0, t1	# .. e1 : bits set iff byte == zero
+	cmpbge	zero, t2, t3	# e0    : bits set iff byte == c
+	beq	t1, $loop	# .. e1 : if we havnt seen a null, loop
+
+	/* Mask out character matches after terminator */
+$eos:
+	negq	t1, t4		# e0    : isolate first null byte match
+	and	t1, t4, t4	# e1    :
+	subq	t4, 1, t5	# e0    : build a mask of the bytes up to...
+	or	t4, t5, t4	# e1    : ... and including the null
+
+	and	t3, t4, t3	# e0    : mask out char matches after null
+	cmovne	t3, t3, t8	# .. e1 : save it, if match found
+	cmovne	t3, v0, t6	# e0    :
+
+	/* Locate the address of the last matched character */
+
+	/* Retain the early exit for the ev4 -- the ev5 mispredict penalty
+	   is 5 cycles -- the same as just falling through.  */
+	beq	t8, $retnull	# .. e1 :
+
+	and	t8, 0xf0, t2	# e0    : binary search for the high bit set
+	cmovne	t2, t2, t8	# .. e1 (zdb)
+	cmovne	t2, 4, t2	# e0    :
+	and	t8, 0xcc, t1	# .. e1 :
+	cmovne	t1, t1, t8	# e0    :
+	cmovne	t1, 2, t1	# .. e1 :
+	and	t8, 0xaa, t0	# e0    :
+	cmovne	t0, 1, t0	# .. e1 (zdb)
+	addq	t2, t1, t1	# e0    :
+	addq	t6, t0, v0	# .. e1 : add our aligned base ptr to the mix
+	addq	v0, t1, v0	# e0    :
+	ret			# .. e1 :
+
+$retnull:
+	mov	zero, v0	# e0    :
+	ret			# .. e1 :
+
+	.end strrchr

arch/alpha/lib/stxcpy.S

@@ -0,0 +1,289 @@
+/*
+ * arch/alpha/lib/stxcpy.S
+ * Contributed by Richard Henderson (rth@tamu.edu)
+ *
+ * Copy a null-terminated string from SRC to DST.
+ *
+ * This is an internal routine used by strcpy, stpcpy, and strcat.
+ * As such, it uses special linkage conventions to make implementation
+ * of these public functions more efficient.
+ *
+ * On input:
+ *	t9 = return address
+ *	a0 = DST
+ *	a1 = SRC
+ *
+ * On output:
+ *	t12 = bitmask (with one bit set) indicating the last byte written
+ *	a0  = unaligned address of the last *word* written
+ *
+ * Furthermore, v0, a3-a5, t11, and t12 are untouched.
+ */
+
+#include <asm/regdef.h>
+
+	.set noat
+	.set noreorder
+
+	.text
+
+/* There is a problem with either gdb (as of 4.16) or gas (as of 2.7) that
+   doesn't like putting the entry point for a procedure somewhere in the
+   middle of the procedure descriptor.  Work around this by putting the
+   aligned copy in its own procedure descriptor */
+
+	.ent stxcpy_aligned
+	.align 3
+stxcpy_aligned:
+	.frame sp, 0, t9
+	.prologue 0
+
+	/* On entry to this basic block:
+	   t0 == the first destination word for masking back in
+	   t1 == the first source word.  */
+
+	/* Create the 1st output word and detect 0's in the 1st input word.  */
+	lda	t2, -1		# e1    : build a mask against false zero
+	mskqh	t2, a1, t2	# e0    :   detection in the src word
+	mskqh	t1, a1, t3	# e0    :
+	ornot	t1, t2, t2	# .. e1 :
+	mskql	t0, a1, t0	# e0    : assemble the first output word
+	cmpbge	zero, t2, t8	# .. e1 : bits set iff null found
+	or	t0, t3, t1	# e0    :
+	bne	t8, $a_eos	# .. e1 :
+
+	/* On entry to this basic block:
+	   t0 == the first destination word for masking back in
+	   t1 == a source word not containing a null.  */
+
+$a_loop:
+	stq_u	t1, 0(a0)	# e0    :
+	addq	a0, 8, a0	# .. e1 :
+	ldq_u	t1, 0(a1)	# e0    :
+	addq	a1, 8, a1	# .. e1 :
+	cmpbge	zero, t1, t8	# e0 (stall)
+	beq	t8, $a_loop	# .. e1 (zdb)
+
+	/* Take care of the final (partial) word store.
+	   On entry to this basic block we have:
+	   t1 == the source word containing the null
+	   t8 == the cmpbge mask that found it.  */
+$a_eos:
+	negq	t8, t6		# e0    : find low bit set
+	and	t8, t6, t12	# e1 (stall)
+
+	/* For the sake of the cache, don't read a destination word
+	   if we're not going to need it.  */
+	and	t12, 0x80, t6	# e0    :
+	bne	t6, 1f		# .. e1 (zdb)
+
+	/* We're doing a partial word store and so need to combine
+	   our source and original destination words.  */
+	ldq_u	t0, 0(a0)	# e0    :
+	subq	t12, 1, t6	# .. e1 :
+	zapnot	t1, t6, t1	# e0    : clear src bytes >= null
+	or	t12, t6, t8	# .. e1 :
+	zap	t0, t8, t0	# e0    : clear dst bytes <= null
+	or	t0, t1, t1	# e1    :
+
+1:	stq_u	t1, 0(a0)	# e0    :
+	ret	(t9)		# .. e1 :
+
+	.end stxcpy_aligned
+
+	.align 3
+	.ent __stxcpy
+	.globl __stxcpy
+__stxcpy:
+	.frame sp, 0, t9
+	.prologue 0
+
+	/* Are source and destination co-aligned?  */
+	xor	a0, a1, t0	# e0    :
+	unop			#       :
+	and	t0, 7, t0	# e0    :
+	bne	t0, $unaligned	# .. e1 :
+
+	/* We are co-aligned; take care of a partial first word.  */
+	ldq_u	t1, 0(a1)	# e0    : load first src word
+	and	a0, 7, t0	# .. e1 : take care not to load a word ...
+	addq	a1, 8, a1		# e0    :
+	beq	t0, stxcpy_aligned	# .. e1 : ... if we wont need it
+	ldq_u	t0, 0(a0)	# e0    :
+	br	stxcpy_aligned	# .. e1 :
+
+
+/* The source and destination are not co-aligned.  Align the destination
+   and cope.  We have to be very careful about not reading too much and
+   causing a SEGV.  */
+
+	.align 3
+$u_head:
+	/* We know just enough now to be able to assemble the first
+	   full source word.  We can still find a zero at the end of it
+	   that prevents us from outputting the whole thing.
+
+	   On entry to this basic block:
+	   t0 == the first dest word, for masking back in, if needed else 0
+	   t1 == the low bits of the first source word
+	   t6 == bytemask that is -1 in dest word bytes */
+
+	ldq_u	t2, 8(a1)	# e0    :
+	addq	a1, 8, a1	# .. e1 :
+
+	extql	t1, a1, t1	# e0    :
+	extqh	t2, a1, t4	# e0    :
+	mskql	t0, a0, t0	# e0    :
+	or	t1, t4, t1	# .. e1 :
+	mskqh	t1, a0, t1	# e0    :
+	or	t0, t1, t1	# e1    :
+
+	or	t1, t6, t6	# e0    :
+	cmpbge	zero, t6, t8	# .. e1 :
+	lda	t6, -1		# e0    : for masking just below
+	bne	t8, $u_final	# .. e1 :
+
+	mskql	t6, a1, t6		# e0    : mask out the bits we have
+	or	t6, t2, t2		# e1    :   already extracted before
+	cmpbge	zero, t2, t8		# e0    :   testing eos
+	bne	t8, $u_late_head_exit	# .. e1 (zdb)
+
+	/* Finally, we've got all the stupid leading edge cases taken care
+	   of and we can set up to enter the main loop.  */
+
+	stq_u	t1, 0(a0)	# e0    : store first output word
+	addq	a0, 8, a0	# .. e1 :
+	extql	t2, a1, t0	# e0    : position ho-bits of lo word
+	ldq_u	t2, 8(a1)	# .. e1 : read next high-order source word
+	addq	a1, 8, a1	# e0    :
+	cmpbge	zero, t2, t8	# .. e1 :
+	nop			# e0    :
+	bne	t8, $u_eos	# .. e1 :
+
+	/* Unaligned copy main loop.  In order to avoid reading too much,
+	   the loop is structured to detect zeros in aligned source words.
+	   This has, unfortunately, effectively pulled half of a loop
+	   iteration out into the head and half into the tail, but it does
+	   prevent nastiness from accumulating in the very thing we want
+	   to run as fast as possible.
+
+	   On entry to this basic block:
+	   t0 == the shifted high-order bits from the previous source word
+	   t2 == the unshifted current source word
+
+	   We further know that t2 does not contain a null terminator.  */
+
+	.align 3
+$u_loop:
+	extqh	t2, a1, t1	# e0    : extract high bits for current word
+	addq	a1, 8, a1	# .. e1 :
+	extql	t2, a1, t3	# e0    : extract low bits for next time
+	addq	a0, 8, a0	# .. e1 :
+	or	t0, t1, t1	# e0    : current dst word now complete
+	ldq_u	t2, 0(a1)	# .. e1 : load high word for next time
+	stq_u	t1, -8(a0)	# e0    : save the current word
+	mov	t3, t0		# .. e1 :
+	cmpbge	zero, t2, t8	# e0    : test new word for eos
+	beq	t8, $u_loop	# .. e1 :
+
+	/* We've found a zero somewhere in the source word we just read.
+	   If it resides in the lower half, we have one (probably partial)
+	   word to write out, and if it resides in the upper half, we
+	   have one full and one partial word left to write out.
+
+	   On entry to this basic block:
+	   t0 == the shifted high-order bits from the previous source word
+	   t2 == the unshifted current source word.  */
+$u_eos:
+	extqh	t2, a1, t1	# e0    :
+	or	t0, t1, t1	# e1    : first (partial) source word complete
+
+	cmpbge	zero, t1, t8	# e0    : is the null in this first bit?
+	bne	t8, $u_final	# .. e1 (zdb)
+
+$u_late_head_exit:
+	stq_u	t1, 0(a0)	# e0    : the null was in the high-order bits
+	addq	a0, 8, a0	# .. e1 :
+	extql	t2, a1, t1	# e0    :
+	cmpbge	zero, t1, t8	# .. e1 :
+
+	/* Take care of a final (probably partial) result word.
+	   On entry to this basic block:
+	   t1 == assembled source word
+	   t8 == cmpbge mask that found the null.  */
+$u_final:
+	negq	t8, t6		# e0    : isolate low bit set
+	and	t6, t8, t12	# e1    :
+
+	and	t12, 0x80, t6	# e0    : avoid dest word load if we can
+	bne	t6, 1f		# .. e1 (zdb)
+
+	ldq_u	t0, 0(a0)	# e0    :
+	subq	t12, 1, t6	# .. e1 :
+	or	t6, t12, t8	# e0    :
+	zapnot	t1, t6, t1	# .. e1 : kill source bytes >= null
+	zap	t0, t8, t0	# e0    : kill dest bytes <= null
+	or	t0, t1, t1	# e1    :
+
+1:	stq_u	t1, 0(a0)	# e0    :
+	ret	(t9)		# .. e1 :
+
+	/* Unaligned copy entry point.  */
+	.align 3
+$unaligned:
+
+	ldq_u	t1, 0(a1)	# e0    : load first source word
+
+	and	a0, 7, t4	# .. e1 : find dest misalignment
+	and	a1, 7, t5	# e0    : find src misalignment
+
+	/* Conditionally load the first destination word and a bytemask
+	   with 0xff indicating that the destination byte is sacrosanct.  */
+
+	mov	zero, t0	# .. e1 :
+	mov	zero, t6	# e0    :
+	beq	t4, 1f		# .. e1 :
+	ldq_u	t0, 0(a0)	# e0    :
+	lda	t6, -1		# .. e1 :
+	mskql	t6, a0, t6	# e0    :
+1:
+	subq	a1, t4, a1	# .. e1 : sub dest misalignment from src addr
+
+	/* If source misalignment is larger than dest misalignment, we need
+	   extra startup checks to avoid SEGV.  */
+
+	cmplt	t4, t5, t12	# e0    :
+	beq	t12, $u_head	# .. e1 (zdb)
+
+	lda	t2, -1		# e1    : mask out leading garbage in source
+	mskqh	t2, t5, t2	# e0    :
+	nop			# e0    :
+	ornot	t1, t2, t3	# .. e1 :
+	cmpbge	zero, t3, t8	# e0    : is there a zero?
+	beq	t8, $u_head	# .. e1 (zdb)
+
+	/* At this point we've found a zero in the first partial word of
+	   the source.  We need to isolate the valid source data and mask
+	   it into the original destination data.  (Incidentally, we know
+	   that we'll need at least one byte of that original dest word.) */
+
+	ldq_u	t0, 0(a0)	# e0    :
+
+	negq	t8, t6		# .. e1 : build bitmask of bytes <= zero
+	and	t6, t8, t12	# e0    :
+	and	a1, 7, t5	# .. e1 :
+	subq	t12, 1, t6	# e0    :
+	or	t6, t12, t8	# e1    :
+	srl	t12, t5, t12	# e0    : adjust final null return value
+
+	zapnot	t2, t8, t2	# .. e1 : prepare source word; mirror changes
+	and	t1, t2, t1	# e1    : to source validity mask
+	extql	t2, a1, t2	# .. e0 :
+	extql	t1, a1, t1	# e0    :
+
+	andnot	t0, t2, t0	# .. e1 : zero place for source to reside
+	or	t0, t1, t1	# e1    : and put it there
+	stq_u	t1, 0(a0)	# .. e0 :
+	ret	(t9)		# e1    :
+
+	.end __stxcpy

arch/alpha/lib/stxncpy.S

@@ -0,0 +1,345 @@
+/*
+ * arch/alpha/lib/stxncpy.S
+ * Contributed by Richard Henderson (rth@tamu.edu)
+ *
+ * Copy no more than COUNT bytes of the null-terminated string from
+ * SRC to DST.
+ *
+ * This is an internal routine used by strncpy, stpncpy, and strncat.
+ * As such, it uses special linkage conventions to make implementation
+ * of these public functions more efficient.
+ *
+ * On input:
+ *	t9 = return address
+ *	a0 = DST
+ *	a1 = SRC
+ *	a2 = COUNT
+ *
+ * Furthermore, COUNT may not be zero.
+ *
+ * On output:
+ *	t0  = last word written
+ *	t10 = bitmask (with one bit set) indicating the byte position of
+ *	      the end of the range specified by COUNT
+ *	t12 = bitmask (with one bit set) indicating the last byte written
+ *	a0  = unaligned address of the last *word* written
+ *	a2  = the number of full words left in COUNT
+ *
+ * Furthermore, v0, a3-a5, t11, and $at are untouched.
+ */
+
+#include <asm/regdef.h>
+
+	.set noat
+	.set noreorder
+
+	.text
+
+/* There is a problem with either gdb (as of 4.16) or gas (as of 2.7) that
+   doesn't like putting the entry point for a procedure somewhere in the
+   middle of the procedure descriptor.  Work around this by putting the
+   aligned copy in its own procedure descriptor */
+
+	.ent stxncpy_aligned
+	.align 3
+stxncpy_aligned:
+	.frame sp, 0, t9, 0
+	.prologue 0
+
+	/* On entry to this basic block:
+	   t0 == the first destination word for masking back in
+	   t1 == the first source word.  */
+
+	/* Create the 1st output word and detect 0's in the 1st input word.  */
+	lda	t2, -1		# e1    : build a mask against false zero
+	mskqh	t2, a1, t2	# e0    :   detection in the src word
+	mskqh	t1, a1, t3	# e0    :
+	ornot	t1, t2, t2	# .. e1 :
+	mskql	t0, a1, t0	# e0    : assemble the first output word
+	cmpbge	zero, t2, t8	# .. e1 : bits set iff null found
+	or	t0, t3, t0	# e0    :
+	beq	a2, $a_eoc	# .. e1 :
+	bne	t8, $a_eos	# .. e1 :
+
+	/* On entry to this basic block:
+	   t0 == a source word not containing a null.  */
+
+$a_loop:
+	stq_u	t0, 0(a0)	# e0    :
+	addq	a0, 8, a0	# .. e1 :
+	ldq_u	t0, 0(a1)	# e0    :
+	addq	a1, 8, a1	# .. e1 :
+	subq	a2, 1, a2	# e0    :
+	cmpbge	zero, t0, t8	# .. e1 (stall)
+	beq	a2, $a_eoc      # e1    :
+	beq	t8, $a_loop	# e1    :
+
+	/* Take care of the final (partial) word store.  At this point
+	   the end-of-count bit is set in t8 iff it applies.
+
+	   On entry to this basic block we have:
+	   t0 == the source word containing the null
+	   t8 == the cmpbge mask that found it.  */
+
+$a_eos:
+	negq	t8, t12		# e0    : find low bit set
+	and	t8, t12, t12	# e1 (stall)
+
+	/* For the sake of the cache, don't read a destination word
+	   if we're not going to need it.  */
+	and	t12, 0x80, t6	# e0    :
+	bne	t6, 1f		# .. e1 (zdb)
+
+	/* We're doing a partial word store and so need to combine
+	   our source and original destination words.  */
+	ldq_u	t1, 0(a0)	# e0    :
+	subq	t12, 1, t6	# .. e1 :
+	or	t12, t6, t8	# e0    :
+	unop			#
+	zapnot	t0, t8, t0	# e0    : clear src bytes > null
+	zap	t1, t8, t1	# .. e1 : clear dst bytes <= null
+	or	t0, t1, t0	# e1    :
+
+1:	stq_u	t0, 0(a0)	# e0    :
+	ret	(t9)		# e1    :
+
+	/* Add the end-of-count bit to the eos detection bitmask.  */
+$a_eoc:
+	or	t10, t8, t8
+	br	$a_eos
+
+	.end stxncpy_aligned
+
+	.align 3
+	.ent __stxncpy
+	.globl __stxncpy
+__stxncpy:
+	.frame sp, 0, t9, 0
+	.prologue 0
+
+	/* Are source and destination co-aligned?  */
+	xor	a0, a1, t1	# e0    :
+	and	a0, 7, t0	# .. e1 : find dest misalignment
+	and	t1, 7, t1	# e0    :
+	addq	a2, t0, a2	# .. e1 : bias count by dest misalignment
+	subq	a2, 1, a2	# e0    :
+	and	a2, 7, t2	# e1    :
+	srl	a2, 3, a2	# e0    : a2 = loop counter = (count - 1)/8
+	addq	zero, 1, t10	# .. e1 :
+	sll	t10, t2, t10	# e0    : t10 = bitmask of last count byte
+	bne	t1, $unaligned	# .. e1 :
+
+	/* We are co-aligned; take care of a partial first word.  */
+
+	ldq_u	t1, 0(a1)	# e0    : load first src word
+	addq	a1, 8, a1	# .. e1 :
+
+	beq	t0, stxncpy_aligned     # avoid loading dest word if not needed
+	ldq_u	t0, 0(a0)	# e0    :
+	br	stxncpy_aligned	# .. e1 :
+
+
+/* The source and destination are not co-aligned.  Align the destination
+   and cope.  We have to be very careful about not reading too much and
+   causing a SEGV.  */
+
+	.align 3
+$u_head:
+	/* We know just enough now to be able to assemble the first
+	   full source word.  We can still find a zero at the end of it
+	   that prevents us from outputting the whole thing.
+
+	   On entry to this basic block:
+	   t0 == the first dest word, unmasked
+	   t1 == the shifted low bits of the first source word
+	   t6 == bytemask that is -1 in dest word bytes */
+
+	ldq_u	t2, 8(a1)	# e0    : load second src word
+	addq	a1, 8, a1	# .. e1 :
+	mskql	t0, a0, t0	# e0    : mask trailing garbage in dst
+	extqh	t2, a1, t4	# e0    :
+	or	t1, t4, t1	# e1    : first aligned src word complete
+	mskqh	t1, a0, t1	# e0    : mask leading garbage in src
+	or	t0, t1, t0	# e0    : first output word complete
+	or	t0, t6, t6	# e1    : mask original data for zero test
+	cmpbge	zero, t6, t8	# e0    :
+	beq	a2, $u_eocfin	# .. e1 :
+	lda	t6, -1		# e0    :
+	bne	t8, $u_final	# .. e1 :
+
+	mskql	t6, a1, t6	# e0    : mask out bits already seen
+	nop			# .. e1 :
+	stq_u	t0, 0(a0)	# e0    : store first output word
+	or      t6, t2, t2	# .. e1 :
+	cmpbge	zero, t2, t8	# e0    : find nulls in second partial
+	addq	a0, 8, a0	# .. e1 :
+	subq	a2, 1, a2	# e0    :
+	bne	t8, $u_late_head_exit	# .. e1 :
+
+	/* Finally, we've got all the stupid leading edge cases taken care
+	   of and we can set up to enter the main loop.  */
+
+	extql	t2, a1, t1	# e0    : position hi-bits of lo word
+	beq	a2, $u_eoc	# .. e1 :
+	ldq_u	t2, 8(a1)	# e0    : read next high-order source word
+	addq	a1, 8, a1	# .. e1 :
+	extqh	t2, a1, t0	# e0    : position lo-bits of hi word (stall)
+	cmpbge	zero, t2, t8	# .. e1 :
+	nop			# e0    :
+	bne	t8, $u_eos	# .. e1 :
+
+	/* Unaligned copy main loop.  In order to avoid reading too much,
+	   the loop is structured to detect zeros in aligned source words.
+	   This has, unfortunately, effectively pulled half of a loop
+	   iteration out into the head and half into the tail, but it does
+	   prevent nastiness from accumulating in the very thing we want
+	   to run as fast as possible.
+
+	   On entry to this basic block:
+	   t0 == the shifted low-order bits from the current source word
+	   t1 == the shifted high-order bits from the previous source word
+	   t2 == the unshifted current source word
+
+	   We further know that t2 does not contain a null terminator.  */
+
+	.align 3
+$u_loop:
+	or	t0, t1, t0	# e0    : current dst word now complete
+	subq	a2, 1, a2	# .. e1 : decrement word count
+	stq_u	t0, 0(a0)	# e0    : save the current word
+	addq	a0, 8, a0	# .. e1 :
+	extql	t2, a1, t1	# e0    : extract high bits for next time
+	beq	a2, $u_eoc	# .. e1 :
+	ldq_u	t2, 8(a1)	# e0    : load high word for next time
+	addq	a1, 8, a1	# .. e1 :
+	nop			# e0    :
+	cmpbge	zero, t2, t8	# e1    : test new word for eos (stall)
+	extqh	t2, a1, t0	# e0    : extract low bits for current word
+	beq	t8, $u_loop	# .. e1 :
+
+	/* We've found a zero somewhere in the source word we just read.
+	   If it resides in the lower half, we have one (probably partial)
+	   word to write out, and if it resides in the upper half, we
+	   have one full and one partial word left to write out.
+
+	   On entry to this basic block:
+	   t0 == the shifted low-order bits from the current source word
+	   t1 == the shifted high-order bits from the previous source word
+	   t2 == the unshifted current source word.  */
+$u_eos:
+	or	t0, t1, t0	# e0    : first (partial) source word complete
+	nop			# .. e1 :
+	cmpbge	zero, t0, t8	# e0    : is the null in this first bit?
+	bne	t8, $u_final	# .. e1 (zdb)
+
+	stq_u	t0, 0(a0)	# e0    : the null was in the high-order bits
+	addq	a0, 8, a0	# .. e1 :
+	subq	a2, 1, a2	# e1    :
+
+$u_late_head_exit:
+	extql	t2, a1, t0	# .. e0 :
+	cmpbge	zero, t0, t8	# e0    :
+	or	t8, t10, t6	# e1    :
+	cmoveq	a2, t6, t8	# e0    :
+	nop			# .. e1 :
+
+	/* Take care of a final (probably partial) result word.
+	   On entry to this basic block:
+	   t0 == assembled source word
+	   t8 == cmpbge mask that found the null.  */
+$u_final:
+	negq	t8, t6		# e0    : isolate low bit set
+	and	t6, t8, t12	# e1    :
+
+	and	t12, 0x80, t6	# e0    : avoid dest word load if we can
+	bne	t6, 1f		# .. e1 (zdb)
+
+	ldq_u	t1, 0(a0)	# e0    :
+	subq	t12, 1, t6	# .. e1 :
+	or	t6, t12, t8	# e0    :
+	zapnot	t0, t8, t0	# .. e1 : kill source bytes > null
+	zap	t1, t8, t1	# e0    : kill dest bytes <= null
+	or	t0, t1, t0	# e1    :
+
+1:	stq_u	t0, 0(a0)	# e0    :
+	ret	(t9)		# .. e1 :
+
+	/* Got to end-of-count before end of string.  
+	   On entry to this basic block:
+	   t1 == the shifted high-order bits from the previous source word  */
+$u_eoc:
+	and	a1, 7, t6	# e1    :
+	sll	t10, t6, t6	# e0    :
+	and	t6, 0xff, t6	# e0    :
+	bne	t6, 1f		# .. e1 :
+
+	ldq_u	t2, 8(a1)	# e0    : load final src word
+	nop			# .. e1 :
+	extqh	t2, a1, t0	# e0    : extract low bits for last word
+	or	t1, t0, t1	# e1    :
+
+1:	cmpbge	zero, t1, t8
+	mov	t1, t0
+
+$u_eocfin:			# end-of-count, final word
+	or	t10, t8, t8
+	br	$u_final
+
+	/* Unaligned copy entry point.  */
+	.align 3
+$unaligned:
+
+	ldq_u	t1, 0(a1)	# e0    : load first source word
+
+	and	a0, 7, t4	# .. e1 : find dest misalignment
+	and	a1, 7, t5	# e0    : find src misalignment
+
+	/* Conditionally load the first destination word and a bytemask
+	   with 0xff indicating that the destination byte is sacrosanct.  */
+
+	mov	zero, t0	# .. e1 :
+	mov	zero, t6	# e0    :
+	beq	t4, 1f		# .. e1 :
+	ldq_u	t0, 0(a0)	# e0    :
+	lda	t6, -1		# .. e1 :
+	mskql	t6, a0, t6	# e0    :
+	subq	a1, t4, a1	# .. e1 : sub dest misalignment from src addr
+
+	/* If source misalignment is larger than dest misalignment, we need
+	   extra startup checks to avoid SEGV.  */
+
+1:	cmplt	t4, t5, t12	# e1    :
+	extql	t1, a1, t1	# .. e0 : shift src into place
+	lda	t2, -1		# e0    : for creating masks later
+	beq	t12, $u_head	# .. e1 :
+
+	extql	t2, a1, t2	# e0    :
+	cmpbge	zero, t1, t8	# .. e1 : is there a zero?
+	andnot	t2, t6, t2	# e0    : dest mask for a single word copy
+	or	t8, t10, t5	# .. e1 : test for end-of-count too
+	cmpbge	zero, t2, t3	# e0    :
+	cmoveq	a2, t5, t8	# .. e1 :
+	andnot	t8, t3, t8	# e0    :
+	beq	t8, $u_head	# .. e1 (zdb)
+
+	/* At this point we've found a zero in the first partial word of
+	   the source.  We need to isolate the valid source data and mask
+	   it into the original destination data.  (Incidentally, we know
+	   that we'll need at least one byte of that original dest word.) */
+
+	ldq_u	t0, 0(a0)	# e0    :
+	negq	t8, t6		# .. e1 : build bitmask of bytes <= zero
+	mskqh	t1, t4, t1	# e0    :
+	and	t6, t8, t12	# .. e1 :
+	subq	t12, 1, t6	# e0    :
+	or	t6, t12, t8	# e1    :
+
+	zapnot	t2, t8, t2	# e0    : prepare source word; mirror changes
+	zapnot	t1, t8, t1	# .. e1 : to source validity mask
+
+	andnot	t0, t2, t0	# e0    : zero place for source to reside
+	or	t0, t1, t0	# e1    : and put it there
+	stq_u	t0, 0(a0)	# e0    :
+	ret	(t9)		# .. e1 :
+
+	.end __stxncpy

arch/alpha/lib/udelay.c

@@ -0,0 +1,54 @@
+/*
+ * Copyright (C) 1993, 2000 Linus Torvalds
+ *
+ * Delay routines, using a pre-computed "loops_per_jiffy" value.
+ */
+
+#include <linux/module.h>
+#include <linux/sched.h> /* for udelay's use of smp_processor_id */
+#include <asm/param.h>
+#include <asm/smp.h>
+#include <linux/delay.h>
+
+/*
+ * Use only for very small delays (< 1 msec). 
+ *
+ * The active part of our cycle counter is only 32-bits wide, and
+ * we're treating the difference between two marks as signed.  On
+ * a 1GHz box, that's about 2 seconds.
+ */
+
+void
+__delay(int loops)
+{
+	int tmp;
+	__asm__ __volatile__(
+		"	rpcc %0\n"
+		"	addl %1,%0,%1\n"
+		"1:	rpcc %0\n"
+		"	subl %1,%0,%0\n"
+		"	bgt %0,1b"
+		: "=&r" (tmp), "=r" (loops) : "1"(loops));
+}
+
+#ifdef CONFIG_SMP
+#define LPJ	 cpu_data[smp_processor_id()].loops_per_jiffy
+#else
+#define LPJ	 loops_per_jiffy
+#endif
+
+void
+udelay(unsigned long usecs)
+{
+	usecs *= (((unsigned long)HZ << 32) / 1000000) * LPJ;
+	__delay((long)usecs >> 32);
+}
+EXPORT_SYMBOL(udelay);
+
+void
+ndelay(unsigned long nsecs)
+{
+	nsecs *= (((unsigned long)HZ << 32) / 1000000000) * LPJ;
+	__delay((long)nsecs >> 32);
+}
+EXPORT_SYMBOL(ndelay);