x86/kprobes: Fix crash when probing CS CALL instructionsgrafted

When using eBPF to probe CS CALL instructions within a function,
a crash can be triggered.

The eBPF tool probes offset 257 of the __hrtimer_run_queues()
function:

<__hrtimer_run_queues+249>:  nopl   0x0(%rax,%rax,1)
<__hrtimer_run_queues+254>:  mov    %r14,%rdi
<__hrtimer_run_queues+257>:  cs call <__x86_indirect_thunk_r12>
<__hrtimer_run_queues+263>:  mov    %eax,%r12d
<__hrtimer_run_queues+266>:  xchg   %ax,%ax
<__hrtimer_run_queues+268>:  mov    %r13,%rdi

Which triggers this crash:

  BUG: unable to handle page fault for address: 00000000000f41c9
  #PF: supervisor write access in kernel mode
  #PF: error_code(0x0002) - not-present page
  PGD 0 P4D 0
  Oops: 0002 [#1] SMP NOPTI
  CPU: 1 PID: 0 Comm: swapper/1 Kdump: loaded Tainted: P
  RIP: 0010:__hrtimer_run_queues+0x106/0x230

Note that __hrtimer_run_queues+0x106 is __hrtimer_run_queues+262, which is
at the 6th byte of the above CS CALL instruction. Since the CS CALL
instruction occupies 6 bytes, the exception occurred in the middle of that
call instruction.

The root cause is that when using eBPF tools to probe in the middle of a
function, a kprobe with INT3 is used as the underlying implementation.

During single-step emulation of the original CALL instruction,
int3_emulate_call() assumes that the probed CALL instruction is 5 bytes
long. However, the actual CS-prefixed CALL instruction occupies 6 bytes,
so it constructs an incorrect exception return address. When the CPU
returns from the kprobe handler, the next instruction to be executed is at
the address of the last byte of that CS CALL instruction. Coincidentally,
starting from that address, the CPU fetches and decodes a completely
different instruction, which ultimately triggers a kernel crash.

Fix the issue by using the actual instruction length obtained from
the instruction decoder when constructing the exception return
address, rather than relying on the hardcoded CALL_INSN_SIZE macro.

[ mingo: Refined the changelog ]

Fixes: 6256e668b7af ("x86/kprobes: Use int3 instead of debug trap for single-step")
Suggested-by: Masami Hiramatsu (Google) <mhiramat@kernel.org>
Signed-off-by: Jinke Han <jinkehan@didiglobal.com>
Signed-off-by: Ingo Molnar <mingo@kernel.org>
Reviewed-by: Masami Hiramatsu (Google) <mhiramat@kernel.org>
Acked-by: Yafang Shao <laoar.shao@gmail.com>
Acked-by: Borislav Petkov <bp@alien8.de>
Cc: Peter Zijlstra <peterz@infradead.org>
Link: https://patch.msgid.link/20260908073742.GA10517@didi-ThinkCentre-M920t-N000

1 files changed, 516 insertions, 0 deletions

diff --git a/arch/sparc/lib/NGmemcpy.S b/arch/sparc/lib/NGmemcpy.S
new file mode 100644
index 000000000..bbd3ea0a6
--- /dev/null
+++ b/arch/sparc/lib/NGmemcpy.S
@@ -0,0 +1,516 @@
+/* SPDX-License-Identifier: GPL-2.0 */
+/* NGmemcpy.S: Niagara optimized memcpy.
+ *
+ * Copyright (C) 2006, 2007 David S. Miller (davem@davemloft.net)
+ */
+
+#ifdef __KERNEL__
+#include <linux/linkage.h>
+#include <asm/asi.h>
+#include <asm/thread_info.h>
+#define GLOBAL_SPARE	%g7
+#define RESTORE_ASI(TMP)	\
+	wr	%g0, ASI_AIUS, %asi
+#else
+#define GLOBAL_SPARE	%g5
+#define RESTORE_ASI(TMP)	\
+	wr	%g0, ASI_PNF, %asi
+#endif
+
+#ifdef __sparc_v9__
+#define SAVE_AMOUNT	128
+#else
+#define SAVE_AMOUNT	64
+#endif
+
+#ifndef STORE_ASI
+#define STORE_ASI	ASI_BLK_INIT_QUAD_LDD_P
+#endif
+
+#ifndef EX_LD
+#define EX_LD(x,y)	x
+#endif
+
+#ifndef EX_ST
+#define EX_ST(x,y)	x
+#endif
+
+#ifndef LOAD
+#ifndef MEMCPY_DEBUG
+#define LOAD(type,addr,dest)	type [addr], dest
+#else
+#define LOAD(type,addr,dest)	type##a [addr] 0x80, dest
+#endif
+#endif
+
+#ifndef LOAD_TWIN
+#define LOAD_TWIN(addr_reg,dest0,dest1)	\
+	ldda [addr_reg] ASI_BLK_INIT_QUAD_LDD_P, dest0
+#endif
+
+#ifndef STORE
+#define STORE(type,src,addr)	type src, [addr]
+#endif
+
+#ifndef STORE_INIT
+#ifndef SIMULATE_NIAGARA_ON_NON_NIAGARA
+#define STORE_INIT(src,addr)	stxa src, [addr] %asi
+#else
+#define STORE_INIT(src,addr)	stx src, [addr + 0x00]
+#endif
+#endif
+
+#ifndef FUNC_NAME
+#define FUNC_NAME	NGmemcpy
+#endif
+
+#ifndef PREAMBLE
+#define PREAMBLE
+#endif
+
+#ifndef XCC
+#define XCC xcc
+#endif
+
+	.register	%g2,#scratch
+	.register	%g3,#scratch
+
+	.text
+#ifndef EX_RETVAL
+#define EX_RETVAL(x)	x
+__restore_asi:
+	wr	%g0, ASI_AIUS, %asi
+	ret
+	 restore
+ENTRY(NG_ret_i2_plus_i4_plus_1)
+	ba,pt	%xcc, __restore_asi
+	 add	%i2, %i5, %i0
+ENDPROC(NG_ret_i2_plus_i4_plus_1)
+ENTRY(NG_ret_i2_plus_g1)
+	ba,pt	%xcc, __restore_asi
+	 add	%i2, %g1, %i0
+ENDPROC(NG_ret_i2_plus_g1)
+ENTRY(NG_ret_i2_plus_g1_minus_8)
+	sub	%g1, 8, %g1
+	ba,pt	%xcc, __restore_asi
+	 add	%i2, %g1, %i0
+ENDPROC(NG_ret_i2_plus_g1_minus_8)
+ENTRY(NG_ret_i2_plus_g1_minus_16)
+	sub	%g1, 16, %g1
+	ba,pt	%xcc, __restore_asi
+	 add	%i2, %g1, %i0
+ENDPROC(NG_ret_i2_plus_g1_minus_16)
+ENTRY(NG_ret_i2_plus_g1_minus_24)
+	sub	%g1, 24, %g1
+	ba,pt	%xcc, __restore_asi
+	 add	%i2, %g1, %i0
+ENDPROC(NG_ret_i2_plus_g1_minus_24)
+ENTRY(NG_ret_i2_plus_g1_minus_32)
+	sub	%g1, 32, %g1
+	ba,pt	%xcc, __restore_asi
+	 add	%i2, %g1, %i0
+ENDPROC(NG_ret_i2_plus_g1_minus_32)
+ENTRY(NG_ret_i2_plus_g1_minus_40)
+	sub	%g1, 40, %g1
+	ba,pt	%xcc, __restore_asi
+	 add	%i2, %g1, %i0
+ENDPROC(NG_ret_i2_plus_g1_minus_40)
+ENTRY(NG_ret_i2_plus_g1_minus_48)
+	sub	%g1, 48, %g1
+	ba,pt	%xcc, __restore_asi
+	 add	%i2, %g1, %i0
+ENDPROC(NG_ret_i2_plus_g1_minus_48)
+ENTRY(NG_ret_i2_plus_g1_minus_56)
+	sub	%g1, 56, %g1
+	ba,pt	%xcc, __restore_asi
+	 add	%i2, %g1, %i0
+ENDPROC(NG_ret_i2_plus_g1_minus_56)
+ENTRY(NG_ret_i2_plus_i4_plus_16)
+        add     %i4, 16, %i4
+	ba,pt	%xcc, __restore_asi
+	 add	%i2, %i4, %i0
+ENDPROC(NG_ret_i2_plus_i4_plus_16)
+ENTRY(NG_ret_i2_plus_i4_plus_8)
+	add	%i4, 8, %i4
+	ba,pt	%xcc, __restore_asi
+	 add	%i2, %i4, %i0
+ENDPROC(NG_ret_i2_plus_i4_plus_8)
+ENTRY(NG_ret_i2_plus_8)
+	ba,pt	%xcc, __restore_asi
+	 add	%i2, 8, %i0
+ENDPROC(NG_ret_i2_plus_8)
+ENTRY(NG_ret_i2_plus_4)
+	ba,pt	%xcc, __restore_asi
+	 add	%i2, 4, %i0
+ENDPROC(NG_ret_i2_plus_4)
+ENTRY(NG_ret_i2_plus_1)
+	ba,pt	%xcc, __restore_asi
+	 add	%i2, 1, %i0
+ENDPROC(NG_ret_i2_plus_1)
+ENTRY(NG_ret_i2_plus_g1_plus_1)
+	add	%g1, 1, %g1
+	ba,pt	%xcc, __restore_asi
+	 add	%i2, %g1, %i0
+ENDPROC(NG_ret_i2_plus_g1_plus_1)
+ENTRY(NG_ret_i2)
+	ba,pt	%xcc, __restore_asi
+	 mov	%i2, %i0
+ENDPROC(NG_ret_i2)
+ENTRY(NG_ret_i2_and_7_plus_i4)
+	and	%i2, 7, %i2
+	ba,pt	%xcc, __restore_asi
+	 add	%i2, %i4, %i0
+ENDPROC(NG_ret_i2_and_7_plus_i4)
+ENTRY(NG_ret_i2_and_7_plus_i4_plus_8)
+	and	%i2, 7, %i2
+	add	%i4, 8, %i4
+	ba,pt	%xcc, __restore_asi
+	 add	%i2, %i4, %i0
+ENDPROC(NG_ret_i2_and_7_plus_i4)
+#endif
+
+	.align		64
+
+	.globl	FUNC_NAME
+	.type	FUNC_NAME,#function
+FUNC_NAME:	/* %i0=dst, %i1=src, %i2=len */
+	PREAMBLE
+	save		%sp, -SAVE_AMOUNT, %sp
+	srlx		%i2, 31, %g2
+	cmp		%g2, 0
+	tne		%xcc, 5
+	mov		%i0, %o0
+	cmp		%i2, 0
+	be,pn		%XCC, 85f
+	 or		%o0, %i1, %i3
+	cmp		%i2, 16
+	blu,a,pn	%XCC, 80f
+	 or		%i3, %i2, %i3
+
+	/* 2 blocks (128 bytes) is the minimum we can do the block
+	 * copy with.  We need to ensure that we'll iterate at least
+	 * once in the block copy loop.  At worst we'll need to align
+	 * the destination to a 64-byte boundary which can chew up
+	 * to (64 - 1) bytes from the length before we perform the
+	 * block copy loop.
+	 */
+	cmp		%i2, (2 * 64)
+	blu,pt		%XCC, 70f
+	 andcc		%i3, 0x7, %g0
+
+	/* %o0:	dst
+	 * %i1:	src
+	 * %i2:	len  (known to be >= 128)
+	 *
+	 * The block copy loops will use %i4/%i5,%g2/%g3 as
+	 * temporaries while copying the data.
+	 */
+
+	LOAD(prefetch, %i1, #one_read)
+	wr		%g0, STORE_ASI, %asi
+
+	/* Align destination on 64-byte boundary.  */
+	andcc		%o0, (64 - 1), %i4
+	be,pt		%XCC, 2f
+	 sub		%i4, 64, %i4
+	sub		%g0, %i4, %i4	! bytes to align dst
+	sub		%i2, %i4, %i2
+1:	subcc		%i4, 1, %i4
+	EX_LD(LOAD(ldub, %i1, %g1), NG_ret_i2_plus_i4_plus_1)
+	EX_ST(STORE(stb, %g1, %o0), NG_ret_i2_plus_i4_plus_1)
+	add		%i1, 1, %i1
+	bne,pt		%XCC, 1b
+	add		%o0, 1, %o0
+
+	/* If the source is on a 16-byte boundary we can do
+	 * the direct block copy loop.  If it is 8-byte aligned
+	 * we can do the 16-byte loads offset by -8 bytes and the
+	 * init stores offset by one register.
+	 *
+	 * If the source is not even 8-byte aligned, we need to do
+	 * shifting and masking (basically integer faligndata).
+	 *
+	 * The careful bit with init stores is that if we store
+	 * to any part of the cache line we have to store the whole
+	 * cacheline else we can end up with corrupt L2 cache line
+	 * contents.  Since the loop works on 64-bytes of 64-byte
+	 * aligned store data at a time, this is easy to ensure.
+	 */
+2:
+	andcc		%i1, (16 - 1), %i4
+	andn		%i2, (64 - 1), %g1	! block copy loop iterator
+	be,pt		%XCC, 50f
+	 sub		%i2, %g1, %i2		! final sub-block copy bytes
+
+	cmp		%i4, 8
+	be,pt		%XCC, 10f
+	 sub		%i1, %i4, %i1
+
+	/* Neither 8-byte nor 16-byte aligned, shift and mask.  */
+	and		%i4, 0x7, GLOBAL_SPARE
+	sll		GLOBAL_SPARE, 3, GLOBAL_SPARE
+	mov		64, %i5
+	EX_LD(LOAD_TWIN(%i1, %g2, %g3), NG_ret_i2_plus_g1)
+	sub		%i5, GLOBAL_SPARE, %i5
+	mov		16, %o4
+	mov		32, %o5
+	mov		48, %o7
+	mov		64, %i3
+
+	bg,pn	   	%XCC, 9f
+	 nop
+
+#define MIX_THREE_WORDS(WORD1, WORD2, WORD3, PRE_SHIFT, POST_SHIFT, TMP) \
+	sllx		WORD1, POST_SHIFT, WORD1; \
+	srlx		WORD2, PRE_SHIFT, TMP; \
+	sllx		WORD2, POST_SHIFT, WORD2; \
+	or		WORD1, TMP, WORD1; \
+	srlx		WORD3, PRE_SHIFT, TMP; \
+	or		WORD2, TMP, WORD2;
+
+8:	EX_LD(LOAD_TWIN(%i1 + %o4, %o2, %o3), NG_ret_i2_plus_g1)
+	MIX_THREE_WORDS(%g2, %g3, %o2, %i5, GLOBAL_SPARE, %o1)
+	LOAD(prefetch, %i1 + %i3, #one_read)
+
+	EX_ST(STORE_INIT(%g2, %o0 + 0x00), NG_ret_i2_plus_g1)
+	EX_ST(STORE_INIT(%g3, %o0 + 0x08), NG_ret_i2_plus_g1_minus_8)
+
+	EX_LD(LOAD_TWIN(%i1 + %o5, %g2, %g3), NG_ret_i2_plus_g1_minus_16)
+	MIX_THREE_WORDS(%o2, %o3, %g2, %i5, GLOBAL_SPARE, %o1)
+
+	EX_ST(STORE_INIT(%o2, %o0 + 0x10), NG_ret_i2_plus_g1_minus_16)
+	EX_ST(STORE_INIT(%o3, %o0 + 0x18), NG_ret_i2_plus_g1_minus_24)
+
+	EX_LD(LOAD_TWIN(%i1 + %o7, %o2, %o3), NG_ret_i2_plus_g1_minus_32)
+	MIX_THREE_WORDS(%g2, %g3, %o2, %i5, GLOBAL_SPARE, %o1)
+
+	EX_ST(STORE_INIT(%g2, %o0 + 0x20), NG_ret_i2_plus_g1_minus_32)
+	EX_ST(STORE_INIT(%g3, %o0 + 0x28), NG_ret_i2_plus_g1_minus_40)
+
+	EX_LD(LOAD_TWIN(%i1 + %i3, %g2, %g3), NG_ret_i2_plus_g1_minus_48)
+	add		%i1, 64, %i1
+	MIX_THREE_WORDS(%o2, %o3, %g2, %i5, GLOBAL_SPARE, %o1)
+
+	EX_ST(STORE_INIT(%o2, %o0 + 0x30), NG_ret_i2_plus_g1_minus_48)
+	EX_ST(STORE_INIT(%o3, %o0 + 0x38), NG_ret_i2_plus_g1_minus_56)
+
+	subcc		%g1, 64, %g1
+	bne,pt		%XCC, 8b
+	 add		%o0, 64, %o0
+
+	ba,pt		%XCC, 60f
+	 add		%i1, %i4, %i1
+
+9:	EX_LD(LOAD_TWIN(%i1 + %o4, %o2, %o3), NG_ret_i2_plus_g1)
+	MIX_THREE_WORDS(%g3, %o2, %o3, %i5, GLOBAL_SPARE, %o1)
+	LOAD(prefetch, %i1 + %i3, #one_read)
+
+	EX_ST(STORE_INIT(%g3, %o0 + 0x00), NG_ret_i2_plus_g1)
+	EX_ST(STORE_INIT(%o2, %o0 + 0x08), NG_ret_i2_plus_g1_minus_8)
+
+	EX_LD(LOAD_TWIN(%i1 + %o5, %g2, %g3), NG_ret_i2_plus_g1_minus_16)
+	MIX_THREE_WORDS(%o3, %g2, %g3, %i5, GLOBAL_SPARE, %o1)
+
+	EX_ST(STORE_INIT(%o3, %o0 + 0x10), NG_ret_i2_plus_g1_minus_16)
+	EX_ST(STORE_INIT(%g2, %o0 + 0x18), NG_ret_i2_plus_g1_minus_24)
+
+	EX_LD(LOAD_TWIN(%i1 + %o7, %o2, %o3), NG_ret_i2_plus_g1_minus_32)
+	MIX_THREE_WORDS(%g3, %o2, %o3, %i5, GLOBAL_SPARE, %o1)
+
+	EX_ST(STORE_INIT(%g3, %o0 + 0x20), NG_ret_i2_plus_g1_minus_32)
+	EX_ST(STORE_INIT(%o2, %o0 + 0x28), NG_ret_i2_plus_g1_minus_40)
+
+	EX_LD(LOAD_TWIN(%i1 + %i3, %g2, %g3), NG_ret_i2_plus_g1_minus_48)
+	add		%i1, 64, %i1
+	MIX_THREE_WORDS(%o3, %g2, %g3, %i5, GLOBAL_SPARE, %o1)
+
+	EX_ST(STORE_INIT(%o3, %o0 + 0x30), NG_ret_i2_plus_g1_minus_48)
+	EX_ST(STORE_INIT(%g2, %o0 + 0x38), NG_ret_i2_plus_g1_minus_56)
+
+	subcc		%g1, 64, %g1
+	bne,pt		%XCC, 9b
+	 add		%o0, 64, %o0
+
+	ba,pt		%XCC, 60f
+	 add		%i1, %i4, %i1
+
+10:	/* Destination is 64-byte aligned, source was only 8-byte
+	 * aligned but it has been subtracted by 8 and we perform
+	 * one twin load ahead, then add 8 back into source when
+	 * we finish the loop.
+	 */
+	EX_LD(LOAD_TWIN(%i1, %o4, %o5), NG_ret_i2_plus_g1)
+	mov	16, %o7
+	mov	32, %g2
+	mov	48, %g3
+	mov	64, %o1
+1:	EX_LD(LOAD_TWIN(%i1 + %o7, %o2, %o3), NG_ret_i2_plus_g1)
+	LOAD(prefetch, %i1 + %o1, #one_read)
+	EX_ST(STORE_INIT(%o5, %o0 + 0x00), NG_ret_i2_plus_g1)	! initializes cache line
+	EX_ST(STORE_INIT(%o2, %o0 + 0x08), NG_ret_i2_plus_g1_minus_8)
+	EX_LD(LOAD_TWIN(%i1 + %g2, %o4, %o5), NG_ret_i2_plus_g1_minus_16)
+	EX_ST(STORE_INIT(%o3, %o0 + 0x10), NG_ret_i2_plus_g1_minus_16)
+	EX_ST(STORE_INIT(%o4, %o0 + 0x18), NG_ret_i2_plus_g1_minus_24)
+	EX_LD(LOAD_TWIN(%i1 + %g3, %o2, %o3), NG_ret_i2_plus_g1_minus_32)
+	EX_ST(STORE_INIT(%o5, %o0 + 0x20), NG_ret_i2_plus_g1_minus_32)
+	EX_ST(STORE_INIT(%o2, %o0 + 0x28), NG_ret_i2_plus_g1_minus_40)
+	EX_LD(LOAD_TWIN(%i1 + %o1, %o4, %o5), NG_ret_i2_plus_g1_minus_48)
+	add		%i1, 64, %i1
+	EX_ST(STORE_INIT(%o3, %o0 + 0x30), NG_ret_i2_plus_g1_minus_48)
+	EX_ST(STORE_INIT(%o4, %o0 + 0x38), NG_ret_i2_plus_g1_minus_56)
+	subcc		%g1, 64, %g1
+	bne,pt		%XCC, 1b
+	 add		%o0, 64, %o0
+
+	ba,pt		%XCC, 60f
+	 add		%i1, 0x8, %i1
+
+50:	/* Destination is 64-byte aligned, and source is 16-byte
+	 * aligned.
+	 */
+	mov	16, %o7
+	mov	32, %g2
+	mov	48, %g3
+	mov	64, %o1
+1:	EX_LD(LOAD_TWIN(%i1 + %g0, %o4, %o5), NG_ret_i2_plus_g1)
+	EX_LD(LOAD_TWIN(%i1 + %o7, %o2, %o3), NG_ret_i2_plus_g1)
+	LOAD(prefetch, %i1 + %o1, #one_read)
+	EX_ST(STORE_INIT(%o4, %o0 + 0x00), NG_ret_i2_plus_g1)	! initializes cache line
+	EX_ST(STORE_INIT(%o5, %o0 + 0x08), NG_ret_i2_plus_g1_minus_8)
+	EX_LD(LOAD_TWIN(%i1 + %g2, %o4, %o5), NG_ret_i2_plus_g1_minus_16)
+	EX_ST(STORE_INIT(%o2, %o0 + 0x10), NG_ret_i2_plus_g1_minus_16)
+	EX_ST(STORE_INIT(%o3, %o0 + 0x18), NG_ret_i2_plus_g1_minus_24)
+	EX_LD(LOAD_TWIN(%i1 + %g3, %o2, %o3), NG_ret_i2_plus_g1_minus_32)
+	add	%i1, 64, %i1
+	EX_ST(STORE_INIT(%o4, %o0 + 0x20), NG_ret_i2_plus_g1_minus_32)
+	EX_ST(STORE_INIT(%o5, %o0 + 0x28), NG_ret_i2_plus_g1_minus_40)
+	EX_ST(STORE_INIT(%o2, %o0 + 0x30), NG_ret_i2_plus_g1_minus_48)
+	EX_ST(STORE_INIT(%o3, %o0 + 0x38), NG_ret_i2_plus_g1_minus_56)
+	subcc	%g1, 64, %g1
+	bne,pt	%XCC, 1b
+	 add	%o0, 64, %o0
+	/* fall through */
+
+60:	
+	membar		#Sync
+
+	/* %i2 contains any final bytes still needed to be copied
+	 * over. If anything is left, we copy it one byte at a time.
+	 */
+	RESTORE_ASI(%i3)
+	brz,pt		%i2, 85f
+	 sub		%o0, %i1, %i3
+	ba,a,pt		%XCC, 90f
+	 nop
+
+	.align		64
+70: /* 16 < len <= 64 */
+	bne,pn		%XCC, 75f
+	 sub		%o0, %i1, %i3
+
+72:
+	andn		%i2, 0xf, %i4
+	and		%i2, 0xf, %i2
+1:	subcc		%i4, 0x10, %i4
+	EX_LD(LOAD(ldx, %i1, %o4), NG_ret_i2_plus_i4_plus_16)
+	add		%i1, 0x08, %i1
+	EX_LD(LOAD(ldx, %i1, %g1), NG_ret_i2_plus_i4_plus_16)
+	sub		%i1, 0x08, %i1
+	EX_ST(STORE(stx, %o4, %i1 + %i3), NG_ret_i2_plus_i4_plus_16)
+	add		%i1, 0x8, %i1
+	EX_ST(STORE(stx, %g1, %i1 + %i3), NG_ret_i2_plus_i4_plus_8)
+	bgu,pt		%XCC, 1b
+	 add		%i1, 0x8, %i1
+73:	andcc		%i2, 0x8, %g0
+	be,pt		%XCC, 1f
+	 nop
+	sub		%i2, 0x8, %i2
+	EX_LD(LOAD(ldx, %i1, %o4), NG_ret_i2_plus_8)
+	EX_ST(STORE(stx, %o4, %i1 + %i3), NG_ret_i2_plus_8)
+	add		%i1, 0x8, %i1
+1:	andcc		%i2, 0x4, %g0
+	be,pt		%XCC, 1f
+	 nop
+	sub		%i2, 0x4, %i2
+	EX_LD(LOAD(lduw, %i1, %i5), NG_ret_i2_plus_4)
+	EX_ST(STORE(stw, %i5, %i1 + %i3), NG_ret_i2_plus_4)
+	add		%i1, 0x4, %i1
+1:	cmp		%i2, 0
+	be,pt		%XCC, 85f
+	 nop
+	ba,pt		%xcc, 90f
+	 nop
+
+75:
+	andcc		%o0, 0x7, %g1
+	sub		%g1, 0x8, %g1
+	be,pn		%icc, 2f
+	 sub		%g0, %g1, %g1
+	sub		%i2, %g1, %i2
+
+1:	subcc		%g1, 1, %g1
+	EX_LD(LOAD(ldub, %i1, %i5), NG_ret_i2_plus_g1_plus_1)
+	EX_ST(STORE(stb, %i5, %i1 + %i3), NG_ret_i2_plus_g1_plus_1)
+	bgu,pt		%icc, 1b
+	 add		%i1, 1, %i1
+
+2:	add		%i1, %i3, %o0
+	andcc		%i1, 0x7, %g1
+	bne,pt		%icc, 8f
+	 sll		%g1, 3, %g1
+
+	cmp		%i2, 16
+	bgeu,pt		%icc, 72b
+	 nop
+	ba,a,pt		%xcc, 73b
+
+8:	mov		64, %i3
+	andn		%i1, 0x7, %i1
+	EX_LD(LOAD(ldx, %i1, %g2), NG_ret_i2)
+	sub		%i3, %g1, %i3
+	andn		%i2, 0x7, %i4
+	sllx		%g2, %g1, %g2
+1:	add		%i1, 0x8, %i1
+	EX_LD(LOAD(ldx, %i1, %g3), NG_ret_i2_and_7_plus_i4)
+	subcc		%i4, 0x8, %i4
+	srlx		%g3, %i3, %i5
+	or		%i5, %g2, %i5
+	EX_ST(STORE(stx, %i5, %o0), NG_ret_i2_and_7_plus_i4_plus_8)
+	add		%o0, 0x8, %o0
+	bgu,pt		%icc, 1b
+	 sllx		%g3, %g1, %g2
+
+	srl		%g1, 3, %g1
+	andcc		%i2, 0x7, %i2
+	be,pn		%icc, 85f
+	 add		%i1, %g1, %i1
+	ba,pt		%xcc, 90f
+	 sub		%o0, %i1, %i3
+
+	.align		64
+80: /* 0 < len <= 16 */
+	andcc		%i3, 0x3, %g0
+	bne,pn		%XCC, 90f
+	 sub		%o0, %i1, %i3
+
+1:
+	subcc		%i2, 4, %i2
+	EX_LD(LOAD(lduw, %i1, %g1), NG_ret_i2_plus_4)
+	EX_ST(STORE(stw, %g1, %i1 + %i3), NG_ret_i2_plus_4)
+	bgu,pt		%XCC, 1b
+	 add		%i1, 4, %i1
+
+85:	ret
+	 restore	EX_RETVAL(%i0), %g0, %o0
+
+	.align		32
+90:
+	subcc		%i2, 1, %i2
+	EX_LD(LOAD(ldub, %i1, %g1), NG_ret_i2_plus_1)
+	EX_ST(STORE(stb, %g1, %i1 + %i3), NG_ret_i2_plus_1)
+	bgu,pt		%XCC, 90b
+	 add		%i1, 1, %i1
+	ret
+	 restore	EX_RETVAL(%i0), %g0, %o0
+
+	.size		FUNC_NAME, .-FUNC_NAME