libc/arm64: add cfi for optimized func

(gdb) bt
    #0  arch_strlen () at machine/arm64/gnu/arch_strlen.S:119
    #1  0x000000004043a2c0 in strchr (s=s@entry=0x408e5e40 "HELLO WORLD", c=c@entry=87) at machine/arch_libc.c:252
    #2  0x0000000040504c6c in strchr (c=87, s=0x408e5e40 "HELLO WORLD") at /home/mi/ssd/dev-system/nuttx/include/string.h:306
    apache#3  test_strchr () at hello_main.c:79
    apache#4  hello_main (argc=argc@entry=1, argv=argv@entry=0x40b5c090) at hello_main.c:132
    apache#5  0x000000004043d958 in nxtask_startup (entrypt=0x40504ba0 <hello_main>, argc=argc@entry=1, argv=argv@entry=0x40b5c090) at sched/task_startup.c:72
    apache#6  0x00000000403f1ab4 in nxtask_start () at task/task_start.c:104
    apache#7  0x0000000000000000 in ?? ()

Signed-off-by: yangao1 <yangao1@xiaomi.com>

Author: yangao1

libs/libc/machine/arm64/arch_memchr.S

@@ -81,6 +81,8 @@
 	.endm
 
 def_fn ARCH_LIBCFUN(memchr)
+	.cfi_sections .debug_frame
+	.cfi_startproc
 	/* Do not dereference srcin if no bytes to compare.  */
 	cbz	cntin, .Lzero_length
 	/*
@@ -174,6 +176,7 @@ def_fn ARCH_LIBCFUN(memchr)
 	mov	result, #0
 	ret
 
+	.cfi_endproc
 	.size	ARCH_LIBCFUN(memchr), . - ARCH_LIBCFUN(memchr)
 
 #endif

libs/libc/machine/arm64/arch_memcmp.S

@@ -95,6 +95,8 @@
         .endm
 
 def_fn ARCH_LIBCFUN(memcmp) p2align=6
+	.cfi_sections .debug_frame
+	.cfi_startproc
 	subs	limit, limit, 8
 	b.lo	L(less8)
 
@@ -197,6 +199,7 @@ L(byte_loop):
 	sub	result, data1w, data2w
 	ret
 
+	.cfi_endproc
 	.size	ARCH_LIBCFUN(memcmp), . - ARCH_LIBCFUN(memcmp)
 
 #endif

libs/libc/machine/arm64/arch_memcpy.S

@@ -112,6 +112,8 @@
 */
 
 def_fn ARCH_LIBCFUN(memcpy) p2align=6
+	.cfi_sections .debug_frame
+	.cfi_startproc
 	prfm	PLDL1KEEP, [src]
 	add	srcend, src, count
 	add	dstend, dstin, count
@@ -233,6 +235,7 @@ L(copy_long):
 	stp	C_l, C_h, [dstend, -16]
 	ret
 
+	.cfi_endproc
 	.size	ARCH_LIBCFUN(memcpy), . - ARCH_LIBCFUN(memcpy)
 
 #endif

libs/libc/machine/arm64/arch_memmove.S

@@ -101,6 +101,8 @@
 */
 
 def_fn ARCH_LIBCFUN(memmove), 6
+	.cfi_sections .debug_frame
+	.cfi_startproc
 	sub	tmp1, dstin, src
 	cmp	count, 96
 	ccmp	tmp1, count, 2, hi
@@ -158,6 +160,7 @@ def_fn ARCH_LIBCFUN(memmove), 6
 	stp	C_l, C_h, [dstin]
 3:	ret
 
+	.cfi_endproc
 	.size	ARCH_LIBCFUN(memmove), . - ARCH_LIBCFUN(memmove)
 
 #endif

libs/libc/machine/arm64/arch_memset.S

@@ -94,6 +94,8 @@
 
 def_fn ARCH_LIBCFUN(memset) p2align=6
 
+	.cfi_sections .debug_frame
+	.cfi_startproc
 	dup	v0.16B, valw
 	add	dstend, dstin, count
 
@@ -243,6 +245,7 @@ L(zva_other):
 	sub	dst, dst, 32		/* Bias dst for tail loop.  */
 	b	L(tail64)
 
+	.cfi_endproc
 	.size	ARCH_LIBCFUN(memset), . - ARCH_LIBCFUN(memset)
 
 #endif

libs/libc/machine/arm64/arch_strchr.S

@@ -85,6 +85,8 @@
 	.endm
 
 def_fn ARCH_LIBCFUN(strchr)
+	.cfi_sections .debug_frame
+	.cfi_startproc
 	/* Magic constant 0x40100401 to allow us to identify which lane
 	   matches the requested byte.  Magic constant 0x80200802 used
 	   similarly for NUL termination.  */
@@ -162,6 +164,7 @@ def_fn ARCH_LIBCFUN(strchr)
 	csel	result, result, xzr, eq
 	ret
 
+	.cfi_endproc
 	.size	ARCH_LIBCFUN(strchr), . - ARCH_LIBCFUN(strchr)
 
 #endif

libs/libc/machine/arm64/arch_strchrnul.S

@@ -81,6 +81,8 @@
 	.endm
 
 def_fn ARCH_LIBCFUN(strchrnul)
+	.cfi_sections .debug_frame
+	.cfi_startproc
 	/* Magic constant 0x40100401 to allow us to identify which lane
 	   matches the termination condition.  */
 	mov	wtmp2, #0x0401
@@ -147,6 +149,7 @@ def_fn ARCH_LIBCFUN(strchrnul)
 	add	result, src, tmp1, lsr #1
 	ret
 
+	.cfi_endproc
 	.size	ARCH_LIBCFUN(strchrnul), . - ARCH_LIBCFUN(strchrnul)
 
 #endif

libs/libc/machine/arm64/arch_strcmp.S

@@ -77,6 +77,8 @@
 
 	/* Start of performance-critical section  -- one 64B cache line.  */
 def_fn ARCH_LIBCFUN(strcmp) p2align=6
+	.cfi_sections .debug_frame
+	.cfi_startproc
 	eor	tmp1, src1, src2
 	mov	zeroones, #REP8_01
 	tst	tmp1, #7
@@ -206,6 +208,7 @@ L(loop_misaligned):
 L(done):
 	sub	result, data1, data2
 	ret
+	.cfi_endproc
 	.size	ARCH_LIBCFUN(strcmp), .-ARCH_LIBCFUN(strcmp)
 
 #endif

libs/libc/machine/arm64/arch_strcpy.S

@@ -108,6 +108,8 @@
 #define MIN_PAGE_SIZE (1 << MIN_PAGE_P2)
 
 def_fn ARCH_LIBCFUN(strcpy) p2align=6
+	.cfi_sections .debug_frame
+	.cfi_startproc
 	/* For moderately short strings, the fastest way to do the copy is to
 	   calculate the length of the string in the same way as strlen, then
 	   essentially do a memcpy of the result.  This avoids the need for
@@ -321,6 +323,7 @@ def_fn ARCH_LIBCFUN(strcpy) p2align=6
 	bic	has_nul2, tmp3, tmp4
 	b	.Lfp_gt8
 
+	.cfi_endproc
 	.size	ARCH_LIBCFUN(strcpy), . - ARCH_LIBCFUN(strcpy)
 
 #endif

libs/libc/machine/arm64/arch_strlen.S

@@ -114,6 +114,8 @@
 	   boundary.  */
 
 def_fn ARCH_LIBCFUN(strlen) p2align=6
+	.cfi_sections .debug_frame
+	.cfi_startproc
 	and	tmp1, srcin, MIN_PAGE_SIZE - 1
 	mov	zeroones, REP8_01
 	cmp	tmp1, MIN_PAGE_SIZE - 16
@@ -243,6 +245,7 @@ L(page_cross):
 	csel	data2, data2, tmp2, eq
 	b	L(page_cross_entry)
 
+	.cfi_endproc
 	.size	ARCH_LIBCFUN(strlen), . - ARCH_LIBCFUN(strlen)
 
 #endif