good.S

/tmp/pack_pad_transpose_1x9xi8_into_2x4x8x4xi8_dispatch_0-53639f.so:	file format elf64-littleaarch64

Disassembly of section .text:

00000000000103cc <.text>:
   103cc: fd 7b be a9  	stp	x29, x30, [sp, #-32]!
   103d0: f4 4f 01 a9  	stp	x20, x19, [sp, #16]
   103d4: fd 03 00 91  	mov	x29, sp
   103d8: 48 04 40 b9  	ldr	w8, [x2, #4]
   103dc: a8 00 00 34  	cbz	w8, 0x103f0 <.text+0x24>
   103e0: f4 4f 41 a9  	ldp	x20, x19, [sp, #16]
   103e4: e0 03 1f 2a  	mov	w0, wzr
   103e8: fd 7b c2 a8  	ldp	x29, x30, [sp], #32
   103ec: c0 03 5f d6  	ret
   103f0: 29 10 40 f9  	ldr	x9, [x1, #32]
   103f4: 0e 01 80 52  	mov	w14, #8
   103f8: 48 00 40 b9  	ldr	w8, [x2]
   103fc: 2d 31 40 a9  	ldp	x13, x12, [x9]
   10400: 0a 7d 7d d3  	ubfiz	x10, x8, #3, #32
   10404: 2b a4 41 29  	ldp	w11, w9, [x1, #12]
   10408: ad 0d 00 91  	add	x13, x13, #3
   1040c: 8c 41 00 91  	add	x12, x12, #16
   10410: 6b f1 7d d3  	lsl	x11, x11, #3
   10414: 02 00 00 14  	b	0x1041c <.text+0x50>
   10418: 49 fe ff 35  	cbnz	w9, 0x103e0 <.text+0x14>
   1041c: e8 ff ff 35  	cbnz	w8, 0x10418 <.text+0x4c>
   10420: ef 03 0a aa  	mov	x15, x10
   10424: 04 00 00 14  	b	0x10434 <.text+0x68>
   10428: ef 01 0b 8b  	add	x15, x15, x11
   1042c: ff 11 00 f1  	cmp	x15, #4
   10430: 4a ff ff 54  	b.ge	0x10418 <.text+0x4c>
   10434: f0 03 1f aa  	mov	x16, xzr
   10438: f1 03 0d aa  	mov	x17, x13
   1043c: e0 03 0c aa  	mov	x0, x12
   10440: 23 01 80 52  	mov	w3, #9
   10444: 07 00 00 14  	b	0x10460 <.text+0x94>
   10448: 10 06 00 91  	add	x16, x16, #1
   1044c: 00 00 02 91  	add	x0, x0, #128
   10450: 31 22 00 91  	add	x17, x17, #8
   10454: e3 03 01 aa  	mov	x3, x1
   10458: 1f 0a 00 f1  	cmp	x16, #2
   1045c: 60 fe ff 54  	b.eq	0x10428 <.text+0x5c>
   10460: 61 20 00 f1  	subs	x1, x3, #8
   10464: e2 03 11 aa  	mov	x2, x17
   10468: 63 b0 8e 9a  	csel	x3, x3, x14, lt
   1046c: e4 03 00 aa  	mov	x4, x0
   10470: 05 02 80 52  	mov	w5, #16
   10474: 2c 00 00 14  	b	0x10524 <.text+0x158>
   10478: 06 3c 01 0e  	umov	w6, v0.b[0]
   1047c: 07 3c 03 0e  	umov	w7, v0.b[1]
   10480: 01 e4 00 6f  	movi	v1.2d, #0000000000000000
   10484: a5 10 00 f1  	subs	x5, x5, #4
   10488: 02 e4 00 6f  	movi	v2.2d, #0000000000000000
   1048c: 42 90 00 91  	add	x2, x2, #36
   10490: 13 3c 0b 0e  	umov	w19, v0.b[5]
   10494: 04 e4 00 6f  	movi	v4.2d, #0000000000000000
   10498: 14 3c 0d 0e  	umov	w20, v0.b[6]
   1049c: c1 1c 02 4e  	mov	v1.h[0], w6
   104a0: e2 1c 02 4e  	mov	v2.h[0], w7
   104a4: 07 3c 09 0e  	umov	w7, v0.b[4]
   104a8: 03 e4 00 6f  	movi	v3.2d, #0000000000000000
   104ac: 06 3c 05 0e  	umov	w6, v0.b[2]
   104b0: 21 28 21 0e  	xtn	v1.8b, v1.8h
   104b4: 42 28 21 0e  	xtn	v2.8b, v2.8h
   104b8: e4 1c 02 4e  	mov	v4.h[0], w7
   104bc: 07 3c 0f 0e  	umov	w7, v0.b[7]
   104c0: 41 04 0c 6e  	mov	v1.s[1], v2.s[0]
   104c4: 02 e4 00 6f  	movi	v2.2d, #0000000000000000
   104c8: 06 e4 00 6f  	movi	v6.2d, #0000000000000000
   104cc: c3 1c 02 4e  	mov	v3.h[0], w6
   104d0: 06 3c 07 0e  	umov	w6, v0.b[3]
   104d4: 62 1e 02 4e  	mov	v2.h[0], w19
   104d8: 80 28 21 0e  	xtn	v0.8b, v4.8h
   104dc: 05 e4 00 6f  	movi	v5.2d, #0000000000000000
   104e0: 86 1e 02 4e  	mov	v6.h[0], w20
   104e4: 63 28 21 0e  	xtn	v3.8b, v3.8h
   104e8: 42 28 21 0e  	xtn	v2.8b, v2.8h
   104ec: c5 1c 02 4e  	mov	v5.h[0], w6
   104f0: 40 04 0c 6e  	mov	v0.s[1], v2.s[0]
   104f4: 02 e4 00 6f  	movi	v2.2d, #0000000000000000
   104f8: c4 28 21 0e  	xtn	v4.8b, v6.8h
   104fc: 61 04 14 6e  	mov	v1.s[2], v3.s[0]
   10500: a3 28 21 0e  	xtn	v3.8b, v5.8h
   10504: 80 04 14 6e  	mov	v0.s[2], v4.s[0]
   10508: e2 1c 02 4e  	mov	v2.h[0], w7
   1050c: 61 04 1c 6e  	mov	v1.s[3], v3.s[0]
   10510: 42 28 21 0e  	xtn	v2.8b, v2.8h
   10514: 40 04 1c 6e  	mov	v0.s[3], v2.s[0]
   10518: 81 80 3f ad  	stp	q1, q0, [x4, #-16]
   1051c: 84 80 00 91  	add	x4, x4, #32
   10520: 40 f9 ff 54  	b.eq	0x10448 <.text+0x7c>
   10524: 00 e4 00 6f  	movi	v0.2d, #0000000000000000
   10528: a6 3c 00 d1  	sub	x6, x5, #15
   1052c: df 04 00 f1  	cmp	x6, #1
   10530: 4b fa ff 54  	b.lt	0x10478 <.text+0xac>
   10534: 46 0c 00 d1  	sub	x6, x2, #3
   10538: 7f 04 00 f1  	cmp	x3, #1
   1053c: c0 00 40 0d  	ld1	{ v0.b }[0], [x6]
   10540: c0 f9 ff 54  	b.eq	0x10478 <.text+0xac>
   10544: 46 08 00 d1  	sub	x6, x2, #2
   10548: 7f 08 00 f1  	cmp	x3, #2
   1054c: c0 04 40 0d  	ld1	{ v0.b }[1], [x6]
   10550: 40 f9 ff 54  	b.eq	0x10478 <.text+0xac>
   10554: 46 04 00 d1  	sub	x6, x2, #1
   10558: 7f 0c 00 f1  	cmp	x3, #3
   1055c: c0 08 40 0d  	ld1	{ v0.b }[2], [x6]
   10560: c0 f8 ff 54  	b.eq	0x10478 <.text+0xac>
   10564: 40 0c 40 0d  	ld1	{ v0.b }[3], [x2]
   10568: 7f 10 00 f1  	cmp	x3, #4
   1056c: 60 f8 ff 54  	b.eq	0x10478 <.text+0xac>
   10570: 46 04 00 91  	add	x6, x2, #1
   10574: 7f 14 00 f1  	cmp	x3, #5
   10578: c0 10 40 0d  	ld1	{ v0.b }[4], [x6]
   1057c: e0 f7 ff 54  	b.eq	0x10478 <.text+0xac>
   10580: 46 08 00 91  	add	x6, x2, #2
   10584: 7f 18 00 f1  	cmp	x3, #6
   10588: c0 14 40 0d  	ld1	{ v0.b }[5], [x6]
   1058c: 60 f7 ff 54  	b.eq	0x10478 <.text+0xac>
   10590: 46 0c 00 91  	add	x6, x2, #3
   10594: 7f 1c 00 f1  	cmp	x3, #7
   10598: c0 18 40 0d  	ld1	{ v0.b }[6], [x6]
   1059c: e0 f6 ff 54  	b.eq	0x10478 <.text+0xac>
   105a0: 46 10 00 91  	add	x6, x2, #4
   105a4: c0 1c 40 0d  	ld1	{ v0.b }[7], [x6]
   105a8: b4 ff ff 17  	b	0x10478 <.text+0xac>

00000000000105ac <iree_hal_executable_library_query>:
   105ac: 1f 20 03 d5  	nop
   105b0: 88 02 08 10  	adr	x8, #65616
   105b4: 1f 0c 00 71  	cmp	w0, #3
   105b8: 00 01 9f 9a  	csel	x0, x8, xzr, eq
   105bc: c0 03 5f d6  	ret