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, 520 insertions, 0 deletions

diff --git a/arch/sparc/lib/U3memcpy.S b/arch/sparc/lib/U3memcpy.S
new file mode 100644
index 000000000..bace3a18f
--- /dev/null
+++ b/arch/sparc/lib/U3memcpy.S
@@ -0,0 +1,520 @@
+/* SPDX-License-Identifier: GPL-2.0 */
+/* U3memcpy.S: UltraSparc-III optimized memcpy.
+ *
+ * Copyright (C) 1999, 2000, 2004 David S. Miller (davem@redhat.com)
+ */
+
+#ifdef __KERNEL__
+#include <linux/linkage.h>
+#include <asm/visasm.h>
+#include <asm/asi.h>
+#define GLOBAL_SPARE	%g7
+#else
+#define ASI_BLK_P 0xf0
+#define FPRS_FEF  0x04
+#ifdef MEMCPY_DEBUG
+#define VISEntryHalf rd %fprs, %o5; wr %g0, FPRS_FEF, %fprs; \
+		     clr %g1; clr %g2; clr %g3; subcc %g0, %g0, %g0;
+#define VISExitHalf and %o5, FPRS_FEF, %o5; wr %o5, 0x0, %fprs
+#else
+#define VISEntryHalf rd %fprs, %o5; wr %g0, FPRS_FEF, %fprs
+#define VISExitHalf and %o5, FPRS_FEF, %o5; wr %o5, 0x0, %fprs
+#endif
+#define GLOBAL_SPARE	%g5
+#endif
+
+#ifndef EX_LD
+#define EX_LD(x,y)	x
+#endif
+#ifndef EX_LD_FP
+#define EX_LD_FP(x,y)	x
+#endif
+
+#ifndef EX_ST
+#define EX_ST(x,y)	x
+#endif
+#ifndef EX_ST_FP
+#define EX_ST_FP(x,y)	x
+#endif
+
+#ifndef LOAD
+#define LOAD(type,addr,dest)	type [addr], dest
+#endif
+
+#ifndef STORE
+#define STORE(type,src,addr)	type src, [addr]
+#endif
+
+#ifndef STORE_BLK
+#define STORE_BLK(src,addr)	stda src, [addr] ASI_BLK_P
+#endif
+
+#ifndef FUNC_NAME
+#define FUNC_NAME	U3memcpy
+#endif
+
+#ifndef PREAMBLE
+#define PREAMBLE
+#endif
+
+#ifndef XCC
+#define XCC xcc
+#endif
+
+	.register	%g2,#scratch
+	.register	%g3,#scratch
+
+	/* Special/non-trivial issues of this code:
+	 *
+	 * 1) %o5 is preserved from VISEntryHalf to VISExitHalf
+	 * 2) Only low 32 FPU registers are used so that only the
+	 *    lower half of the FPU register set is dirtied by this
+	 *    code.  This is especially important in the kernel.
+	 * 3) This code never prefetches cachelines past the end
+	 *    of the source buffer.
+	 */
+
+	.text
+#ifndef EX_RETVAL
+#define EX_RETVAL(x)	x
+__restore_fp:
+	VISExitHalf
+	retl
+	 nop
+ENTRY(U3_retl_o2_plus_g2_plus_g1_plus_1_fp)
+	add	%g1, 1, %g1
+	add	%g2, %g1, %g2
+	ba,pt	%xcc, __restore_fp
+	 add	%o2, %g2, %o0
+ENDPROC(U3_retl_o2_plus_g2_plus_g1_plus_1_fp)
+ENTRY(U3_retl_o2_plus_g2_fp)
+	ba,pt	%xcc, __restore_fp
+	 add	%o2, %g2, %o0
+ENDPROC(U3_retl_o2_plus_g2_fp)
+ENTRY(U3_retl_o2_plus_g2_plus_8_fp)
+	add	%g2, 8, %g2
+	ba,pt	%xcc, __restore_fp
+	 add	%o2, %g2, %o0
+ENDPROC(U3_retl_o2_plus_g2_plus_8_fp)
+ENTRY(U3_retl_o2)
+	retl
+	 mov	%o2, %o0
+ENDPROC(U3_retl_o2)
+ENTRY(U3_retl_o2_plus_1)
+	retl
+	 add	%o2, 1, %o0
+ENDPROC(U3_retl_o2_plus_1)
+ENTRY(U3_retl_o2_plus_4)
+	retl
+	 add	%o2, 4, %o0
+ENDPROC(U3_retl_o2_plus_4)
+ENTRY(U3_retl_o2_plus_8)
+	retl
+	 add	%o2, 8, %o0
+ENDPROC(U3_retl_o2_plus_8)
+ENTRY(U3_retl_o2_plus_g1_plus_1)
+	add	%g1, 1, %g1
+	retl
+	 add	%o2, %g1, %o0
+ENDPROC(U3_retl_o2_plus_g1_plus_1)
+ENTRY(U3_retl_o2_fp)
+	ba,pt	%xcc, __restore_fp
+	 mov	%o2, %o0
+ENDPROC(U3_retl_o2_fp)
+ENTRY(U3_retl_o2_plus_o3_sll_6_plus_0x80_fp)
+	sll	%o3, 6, %o3
+	add	%o3, 0x80, %o3
+	ba,pt	%xcc, __restore_fp
+	 add	%o2, %o3, %o0
+ENDPROC(U3_retl_o2_plus_o3_sll_6_plus_0x80_fp)
+ENTRY(U3_retl_o2_plus_o3_sll_6_plus_0x40_fp)
+	sll	%o3, 6, %o3
+	add	%o3, 0x40, %o3
+	ba,pt	%xcc, __restore_fp
+	 add	%o2, %o3, %o0
+ENDPROC(U3_retl_o2_plus_o3_sll_6_plus_0x40_fp)
+ENTRY(U3_retl_o2_plus_GS_plus_0x10)
+	add	GLOBAL_SPARE, 0x10, GLOBAL_SPARE
+	retl
+	 add	%o2, GLOBAL_SPARE, %o0
+ENDPROC(U3_retl_o2_plus_GS_plus_0x10)
+ENTRY(U3_retl_o2_plus_GS_plus_0x08)
+	add	GLOBAL_SPARE, 0x08, GLOBAL_SPARE
+	retl
+	 add	%o2, GLOBAL_SPARE, %o0
+ENDPROC(U3_retl_o2_plus_GS_plus_0x08)
+ENTRY(U3_retl_o2_and_7_plus_GS)
+	and	%o2, 7, %o2
+	retl
+	 add	%o2, GLOBAL_SPARE, %o0
+ENDPROC(U3_retl_o2_and_7_plus_GS)
+ENTRY(U3_retl_o2_and_7_plus_GS_plus_8)
+	add	GLOBAL_SPARE, 8, GLOBAL_SPARE
+	and	%o2, 7, %o2
+	retl
+	 add	%o2, GLOBAL_SPARE, %o0
+ENDPROC(U3_retl_o2_and_7_plus_GS_plus_8)
+#endif
+
+	.align		64
+
+	/* The cheetah's flexible spine, oversized liver, enlarged heart,
+	 * slender muscular body, and claws make it the swiftest hunter
+	 * in Africa and the fastest animal on land.  Can reach speeds
+	 * of up to 2.4GB per second.
+	 */
+
+	.globl	FUNC_NAME
+	.type	FUNC_NAME,#function
+FUNC_NAME:	/* %o0=dst, %o1=src, %o2=len */
+	srlx		%o2, 31, %g2
+	cmp		%g2, 0
+
+	/* software trap 5 "Range Check" if dst >= 0x80000000 */
+	tne		%xcc, 5
+	PREAMBLE
+	mov		%o0, %o4
+
+	/* if len == 0 */
+	cmp		%o2, 0
+	be,pn		%XCC, end_return
+	 or		%o0, %o1, %o3
+
+	/* if len < 16 */
+	cmp		%o2, 16
+	blu,a,pn	%XCC, less_than_16
+	 or		%o3, %o2, %o3
+
+	/* if len < 192 */
+	cmp		%o2, (3 * 64)
+	blu,pt		%XCC, less_than_192
+	 andcc		%o3, 0x7, %g0
+
+	/* Clobbers o5/g1/g2/g3/g7/icc/xcc.  We must preserve
+	 * o5 from here until we hit VISExitHalf.
+	 */
+	VISEntryHalf
+
+	/* Is 'dst' already aligned on an 64-byte boundary? */
+	andcc		%o0, 0x3f, %g2
+	be,pt		%XCC, 2f
+
+	/* Compute abs((dst & 0x3f) - 0x40) into %g2.  This is the number
+	 * of bytes to copy to make 'dst' 64-byte aligned.  We pre-
+	 * subtract this from 'len'.
+	 */
+	 sub		%o0, %o1, GLOBAL_SPARE
+	sub		%g2, 0x40, %g2
+	sub		%g0, %g2, %g2
+	sub		%o2, %g2, %o2
+	andcc		%g2, 0x7, %g1
+	be,pt		%icc, 2f
+	 and		%g2, 0x38, %g2
+
+1:	subcc		%g1, 0x1, %g1
+	EX_LD_FP(LOAD(ldub, %o1 + 0x00, %o3), U3_retl_o2_plus_g2_plus_g1_plus_1)
+	EX_ST_FP(STORE(stb, %o3, %o1 + GLOBAL_SPARE), U3_retl_o2_plus_g2_plus_g1_plus_1)
+	bgu,pt		%XCC, 1b
+	 add		%o1, 0x1, %o1
+
+	add		%o1, GLOBAL_SPARE, %o0
+
+2:	cmp		%g2, 0x0
+	and		%o1, 0x7, %g1
+	be,pt		%icc, 3f
+	 alignaddr	%o1, %g0, %o1
+
+	EX_LD_FP(LOAD(ldd, %o1, %f4), U3_retl_o2_plus_g2)
+1:	EX_LD_FP(LOAD(ldd, %o1 + 0x8, %f6), U3_retl_o2_plus_g2)
+	add		%o1, 0x8, %o1
+	subcc		%g2, 0x8, %g2
+	faligndata	%f4, %f6, %f0
+	EX_ST_FP(STORE(std, %f0, %o0), U3_retl_o2_plus_g2_plus_8)
+	be,pn		%icc, 3f
+	 add		%o0, 0x8, %o0
+
+	EX_LD_FP(LOAD(ldd, %o1 + 0x8, %f4), U3_retl_o2_plus_g2)
+	add		%o1, 0x8, %o1
+	subcc		%g2, 0x8, %g2
+	faligndata	%f6, %f4, %f2
+	EX_ST_FP(STORE(std, %f2, %o0), U3_retl_o2_plus_g2_plus_8)
+	bne,pt		%icc, 1b
+	 add		%o0, 0x8, %o0
+
+3:	LOAD(prefetch, %o1 + 0x000, #one_read)
+	LOAD(prefetch, %o1 + 0x040, #one_read)
+	andn		%o2, (0x40 - 1), GLOBAL_SPARE
+	LOAD(prefetch, %o1 + 0x080, #one_read)
+	LOAD(prefetch, %o1 + 0x0c0, #one_read)
+	LOAD(prefetch, %o1 + 0x100, #one_read)
+	EX_LD_FP(LOAD(ldd, %o1 + 0x000, %f0), U3_retl_o2)
+	LOAD(prefetch, %o1 + 0x140, #one_read)
+	EX_LD_FP(LOAD(ldd, %o1 + 0x008, %f2), U3_retl_o2)
+	LOAD(prefetch, %o1 + 0x180, #one_read)
+	EX_LD_FP(LOAD(ldd, %o1 + 0x010, %f4), U3_retl_o2)
+	LOAD(prefetch, %o1 + 0x1c0, #one_read)
+	faligndata	%f0, %f2, %f16
+	EX_LD_FP(LOAD(ldd, %o1 + 0x018, %f6), U3_retl_o2)
+	faligndata	%f2, %f4, %f18
+	EX_LD_FP(LOAD(ldd, %o1 + 0x020, %f8), U3_retl_o2)
+	faligndata	%f4, %f6, %f20
+	EX_LD_FP(LOAD(ldd, %o1 + 0x028, %f10), U3_retl_o2)
+	faligndata	%f6, %f8, %f22
+
+	EX_LD_FP(LOAD(ldd, %o1 + 0x030, %f12), U3_retl_o2)
+	faligndata	%f8, %f10, %f24
+	EX_LD_FP(LOAD(ldd, %o1 + 0x038, %f14), U3_retl_o2)
+	faligndata	%f10, %f12, %f26
+	EX_LD_FP(LOAD(ldd, %o1 + 0x040, %f0), U3_retl_o2)
+
+	and		%o2, 0x3f, %o2
+	subcc		GLOBAL_SPARE, 0x80, GLOBAL_SPARE
+	add		%o1, 0x40, %o1
+	bgu,pt		%XCC, 1f
+	 srl		GLOBAL_SPARE, 6, %o3
+	ba,pt		%xcc, 2f
+	 nop
+
+	.align		64
+1:
+	EX_LD_FP(LOAD(ldd, %o1 + 0x008, %f2), U3_retl_o2_plus_o3_sll_6_plus_0x80)
+	faligndata	%f12, %f14, %f28
+	EX_LD_FP(LOAD(ldd, %o1 + 0x010, %f4), U3_retl_o2_plus_o3_sll_6_plus_0x80)
+	faligndata	%f14, %f0, %f30
+	EX_ST_FP(STORE_BLK(%f16, %o0), U3_retl_o2_plus_o3_sll_6_plus_0x80)
+	EX_LD_FP(LOAD(ldd, %o1 + 0x018, %f6), U3_retl_o2_plus_o3_sll_6_plus_0x40)
+	faligndata	%f0, %f2, %f16
+	add		%o0, 0x40, %o0
+
+	EX_LD_FP(LOAD(ldd, %o1 + 0x020, %f8), U3_retl_o2_plus_o3_sll_6_plus_0x40)
+	faligndata	%f2, %f4, %f18
+	EX_LD_FP(LOAD(ldd, %o1 + 0x028, %f10), U3_retl_o2_plus_o3_sll_6_plus_0x40)
+	faligndata	%f4, %f6, %f20
+	EX_LD_FP(LOAD(ldd, %o1 + 0x030, %f12), U3_retl_o2_plus_o3_sll_6_plus_0x40)
+	subcc		%o3, 0x01, %o3
+	faligndata	%f6, %f8, %f22
+	EX_LD_FP(LOAD(ldd, %o1 + 0x038, %f14), U3_retl_o2_plus_o3_sll_6_plus_0x80)
+
+	faligndata	%f8, %f10, %f24
+	EX_LD_FP(LOAD(ldd, %o1 + 0x040, %f0), U3_retl_o2_plus_o3_sll_6_plus_0x80)
+	LOAD(prefetch, %o1 + 0x1c0, #one_read)
+	faligndata	%f10, %f12, %f26
+	bg,pt		%XCC, 1b
+	 add		%o1, 0x40, %o1
+
+	/* Finally we copy the last full 64-byte block. */
+2:
+	EX_LD_FP(LOAD(ldd, %o1 + 0x008, %f2), U3_retl_o2_plus_o3_sll_6_plus_0x80)
+	faligndata	%f12, %f14, %f28
+	EX_LD_FP(LOAD(ldd, %o1 + 0x010, %f4), U3_retl_o2_plus_o3_sll_6_plus_0x80)
+	faligndata	%f14, %f0, %f30
+	EX_ST_FP(STORE_BLK(%f16, %o0), U3_retl_o2_plus_o3_sll_6_plus_0x80)
+	EX_LD_FP(LOAD(ldd, %o1 + 0x018, %f6), U3_retl_o2_plus_o3_sll_6_plus_0x40)
+	faligndata	%f0, %f2, %f16
+	EX_LD_FP(LOAD(ldd, %o1 + 0x020, %f8), U3_retl_o2_plus_o3_sll_6_plus_0x40)
+	faligndata	%f2, %f4, %f18
+	EX_LD_FP(LOAD(ldd, %o1 + 0x028, %f10), U3_retl_o2_plus_o3_sll_6_plus_0x40)
+	faligndata	%f4, %f6, %f20
+	EX_LD_FP(LOAD(ldd, %o1 + 0x030, %f12), U3_retl_o2_plus_o3_sll_6_plus_0x40)
+	faligndata	%f6, %f8, %f22
+	EX_LD_FP(LOAD(ldd, %o1 + 0x038, %f14), U3_retl_o2_plus_o3_sll_6_plus_0x40)
+	faligndata	%f8, %f10, %f24
+	cmp		%g1, 0
+	be,pt		%XCC, 1f
+	 add		%o0, 0x40, %o0
+	EX_LD_FP(LOAD(ldd, %o1 + 0x040, %f0), U3_retl_o2_plus_o3_sll_6_plus_0x40)
+1:	faligndata	%f10, %f12, %f26
+	faligndata	%f12, %f14, %f28
+	faligndata	%f14, %f0, %f30
+	EX_ST_FP(STORE_BLK(%f16, %o0), U3_retl_o2_plus_o3_sll_6_plus_0x40)
+	add		%o0, 0x40, %o0
+	add		%o1, 0x40, %o1
+	membar		#Sync
+
+	/* Now we copy the (len modulo 64) bytes at the end.
+	 * Note how we borrow the %f0 loaded above.
+	 *
+	 * Also notice how this code is careful not to perform a
+	 * load past the end of the src buffer.
+	 */
+	andcc		%o2, 0x38, %g2
+	be,pn		%XCC, 2f
+	 subcc		%g2, 0x8, %g2
+	be,pn		%XCC, 2f
+	 cmp		%g1, 0
+
+	sub		%o2, %g2, %o2
+	be,a,pt		%XCC, 1f
+	 EX_LD_FP(LOAD(ldd, %o1 + 0x00, %f0), U3_retl_o2_plus_g2)
+
+1:	EX_LD_FP(LOAD(ldd, %o1 + 0x08, %f2), U3_retl_o2_plus_g2)
+	add		%o1, 0x8, %o1
+	subcc		%g2, 0x8, %g2
+	faligndata	%f0, %f2, %f8
+	EX_ST_FP(STORE(std, %f8, %o0), U3_retl_o2_plus_g2_plus_8)
+	be,pn		%XCC, 2f
+	 add		%o0, 0x8, %o0
+	EX_LD_FP(LOAD(ldd, %o1 + 0x08, %f0), U3_retl_o2_plus_g2)
+	add		%o1, 0x8, %o1
+	subcc		%g2, 0x8, %g2
+	faligndata	%f2, %f0, %f8
+	EX_ST_FP(STORE(std, %f8, %o0), U3_retl_o2_plus_g2_plus_8)
+	bne,pn		%XCC, 1b
+	 add		%o0, 0x8, %o0
+
+	/* If anything is left, we copy it one byte at a time.
+	 * Note that %g1 is (src & 0x3) saved above before the
+	 * alignaddr was performed.
+	 */
+2:
+	cmp		%o2, 0
+	add		%o1, %g1, %o1
+	VISExitHalf
+	be,pn		%XCC, end_return
+	 sub		%o0, %o1, %o3
+
+	andcc		%g1, 0x7, %g0
+	bne,pn		%icc, 90f
+	 andcc		%o2, 0x8, %g0
+	be,pt		%icc, 1f
+	 nop
+	EX_LD(LOAD(ldx, %o1, %o5), U3_retl_o2)
+	EX_ST(STORE(stx, %o5, %o1 + %o3), U3_retl_o2)
+	add		%o1, 0x8, %o1
+	sub		%o2, 8, %o2
+
+1:	andcc		%o2, 0x4, %g0
+	be,pt		%icc, 1f
+	 nop
+	EX_LD(LOAD(lduw, %o1, %o5), U3_retl_o2)
+	EX_ST(STORE(stw, %o5, %o1 + %o3), U3_retl_o2)
+	add		%o1, 0x4, %o1
+	sub		%o2, 4, %o2
+
+1:	andcc		%o2, 0x2, %g0
+	be,pt		%icc, 1f
+	 nop
+	EX_LD(LOAD(lduh, %o1, %o5), U3_retl_o2)
+	EX_ST(STORE(sth, %o5, %o1 + %o3), U3_retl_o2)
+	add		%o1, 0x2, %o1
+	sub		%o2, 2, %o2
+
+1:	andcc		%o2, 0x1, %g0
+	be,pt		%icc, end_return
+	 nop
+	EX_LD(LOAD(ldub, %o1, %o5), U3_retl_o2)
+	ba,pt		%xcc, end_return
+	 EX_ST(STORE(stb, %o5, %o1 + %o3), U3_retl_o2)
+
+	.align		64
+	/* 16 <= len < 192 */
+less_than_192:
+	bne,pn		%XCC, 75f
+	 sub		%o0, %o1, %o3
+
+72:
+	andn		%o2, 0xf, GLOBAL_SPARE
+	and		%o2, 0xf, %o2
+1:	subcc		GLOBAL_SPARE, 0x10, GLOBAL_SPARE
+	EX_LD(LOAD(ldx, %o1 + 0x00, %o5), U3_retl_o2_plus_GS_plus_0x10)
+	EX_LD(LOAD(ldx, %o1 + 0x08, %g1), U3_retl_o2_plus_GS_plus_0x10)
+	EX_ST(STORE(stx, %o5, %o1 + %o3), U3_retl_o2_plus_GS_plus_0x10)
+	add		%o1, 0x8, %o1
+	EX_ST(STORE(stx, %g1, %o1 + %o3), U3_retl_o2_plus_GS_plus_0x08)
+	bgu,pt		%XCC, 1b
+	 add		%o1, 0x8, %o1
+73:	andcc		%o2, 0x8, %g0
+	be,pt		%XCC, 1f
+	 nop
+	sub		%o2, 0x8, %o2
+	EX_LD(LOAD(ldx, %o1, %o5), U3_retl_o2_plus_8)
+	EX_ST(STORE(stx, %o5, %o1 + %o3), U3_retl_o2_plus_8)
+	add		%o1, 0x8, %o1
+1:	andcc		%o2, 0x4, %g0
+	be,pt		%XCC, 1f
+	 nop
+	sub		%o2, 0x4, %o2
+	EX_LD(LOAD(lduw, %o1, %o5), U3_retl_o2_plus_4)
+	EX_ST(STORE(stw, %o5, %o1 + %o3), U3_retl_o2_plus_4)
+	add		%o1, 0x4, %o1
+1:	cmp		%o2, 0
+	be,pt		%XCC, end_return
+	 nop
+	ba,pt		%xcc, 90f
+	 nop
+
+75:
+	andcc		%o0, 0x7, %g1
+	sub		%g1, 0x8, %g1
+	be,pn		%icc, 2f
+	 sub		%g0, %g1, %g1
+	sub		%o2, %g1, %o2
+
+1:	subcc		%g1, 1, %g1
+	EX_LD(LOAD(ldub, %o1, %o5), U3_retl_o2_plus_g1_plus_1)
+	EX_ST(STORE(stb, %o5, %o1 + %o3), U3_retl_o2_plus_g1_plus_1)
+	bgu,pt		%icc, 1b
+	 add		%o1, 1, %o1
+
+2:	add		%o1, %o3, %o0
+	andcc		%o1, 0x7, %g1
+	bne,pt		%icc, 8f
+	 sll		%g1, 3, %g1
+
+	cmp		%o2, 16
+	bgeu,pt		%icc, 72b
+	 nop
+	ba,a,pt		%xcc, 73b
+
+8:	mov		64, %o3
+	andn		%o1, 0x7, %o1
+	EX_LD(LOAD(ldx, %o1, %g2), U3_retl_o2)
+	sub		%o3, %g1, %o3
+	andn		%o2, 0x7, GLOBAL_SPARE
+	sllx		%g2, %g1, %g2
+1:	EX_LD(LOAD(ldx, %o1 + 0x8, %g3), U3_retl_o2_and_7_plus_GS)
+	subcc		GLOBAL_SPARE, 0x8, GLOBAL_SPARE
+	add		%o1, 0x8, %o1
+	srlx		%g3, %o3, %o5
+	or		%o5, %g2, %o5
+	EX_ST(STORE(stx, %o5, %o0), U3_retl_o2_and_7_plus_GS_plus_8)
+	add		%o0, 0x8, %o0
+	bgu,pt		%icc, 1b
+	 sllx		%g3, %g1, %g2
+
+	srl		%g1, 3, %g1
+	andcc		%o2, 0x7, %o2
+	be,pn		%icc, end_return
+	 add		%o1, %g1, %o1
+	ba,pt		%xcc, 90f
+	 sub		%o0, %o1, %o3
+
+	.align		64
+	/* 0 < len < 16 */
+less_than_16:
+	andcc		%o3, 0x3, %g0
+	bne,pn		%XCC, 90f
+	 sub		%o0, %o1, %o3
+
+1:
+	subcc		%o2, 4, %o2
+	EX_LD(LOAD(lduw, %o1, %g1), U3_retl_o2_plus_4)
+	EX_ST(STORE(stw, %g1, %o1 + %o3), U3_retl_o2_plus_4)
+	bgu,pt		%XCC, 1b
+	 add		%o1, 4, %o1
+
+end_return:
+	retl
+	 mov		EX_RETVAL(%o4), %o0
+
+	.align		32
+90:
+	subcc		%o2, 1, %o2
+	EX_LD(LOAD(ldub, %o1, %g1), U3_retl_o2_plus_1)
+	EX_ST(STORE(stb, %g1, %o1 + %o3), U3_retl_o2_plus_1)
+	bgu,pt		%XCC, 90b
+	 add		%o1, 1, %o1
+	retl
+	 mov		EX_RETVAL(%o4), %o0
+
+	.size		FUNC_NAME, .-FUNC_NAME