Skip to content

Commit d58337e

Browse files
kelindarkelindar
committed
regenerate with omit frame pointer
1 parent ad2a037 commit d58337e

File tree

2 files changed

+76
-130
lines changed

2 files changed

+76
-130
lines changed

‎simd_apple.s‎

Lines changed: 26 additions & 53 deletions
Original file line numberDiff line numberDiff line change
@@ -5,18 +5,16 @@ TEXT ·_and(SB), $0-32
55
MOVD a+0(FP), R0
66
MOVD b+8(FP), R1
77
MOVD n+16(FP), R2
8-
WORD $0xa9bf7bfd // stp x29, x30, [sp, #-16]! ; 16-byte Folded Spill
9-
WORD $0x910003fd // mov x29, sp
108
WORD $0xb40002a2 // cbz x2, LBB0_7
119
WORD $0xf100105f // cmp x2, #4
1210
WORD $0x54000103 // b.lo LBB0_4
1311
WORD $0xd37df048 // lsl x8, x2, #3
1412
WORD $0x8b080029 // add x9, x1, x8
1513
WORD $0xeb00013f // cmp x9, x0
16-
WORD $0x54000229 // b.ls LBB0_8
14+
WORD $0x54000209 // b.ls LBB0_8
1715
WORD $0x8b080008 // add x8, x0, x8
1816
WORD $0xeb01011f // cmp x8, x1
19-
WORD $0x540001c9 // b.ls LBB0_8
17+
WORD $0x540001a9 // b.ls LBB0_8
2018

2119
BB0_4:
2220
WORD $0xd2800008 // mov x8, #0
@@ -36,7 +34,6 @@ BB0_6:
3634
WORD $0x54ffff61 // b.ne LBB0_6
3735

3836
BB0_7:
39-
WORD $0xa8c17bfd // ldp x29, x30, [sp], #16 ; 16-byte Folded Reload
4037
WORD $0xd65f03c0 // ret
4138

4239
BB0_8:
@@ -56,25 +53,23 @@ BB0_9:
5653
WORD $0xf100116b // subs x11, x11, #4
5754
WORD $0x54ffff01 // b.ne LBB0_9
5855
WORD $0xeb02011f // cmp x8, x2
59-
WORD $0x54fffe00 // b.eq LBB0_7
60-
WORD $0x17ffffe5 // b LBB0_5
56+
WORD $0x54fffe20 // b.eq LBB0_7
57+
WORD $0x17ffffe6 // b LBB0_5
6158

6259
TEXT ·_andn(SB), $0-32
6360
MOVD a+0(FP), R0
6461
MOVD b+8(FP), R1
6562
MOVD n+16(FP), R2
66-
WORD $0xa9bf7bfd // stp x29, x30, [sp, #-16]! ; 16-byte Folded Spill
67-
WORD $0x910003fd // mov x29, sp
6863
WORD $0xb40002a2 // cbz x2, LBB1_7
6964
WORD $0xf100105f // cmp x2, #4
7065
WORD $0x54000103 // b.lo LBB1_4
7166
WORD $0xd37df048 // lsl x8, x2, #3
7267
WORD $0x8b080029 // add x9, x1, x8
7368
WORD $0xeb00013f // cmp x9, x0
74-
WORD $0x54000229 // b.ls LBB1_8
69+
WORD $0x54000209 // b.ls LBB1_8
7570
WORD $0x8b080008 // add x8, x0, x8
7671
WORD $0xeb01011f // cmp x8, x1
77-
WORD $0x540001c9 // b.ls LBB1_8
72+
WORD $0x540001a9 // b.ls LBB1_8
7873

7974
BB1_4:
8075
WORD $0xd2800008 // mov x8, #0
@@ -94,7 +89,6 @@ BB1_6:
9489
WORD $0x54ffff61 // b.ne LBB1_6
9590

9691
BB1_7:
97-
WORD $0xa8c17bfd // ldp x29, x30, [sp], #16 ; 16-byte Folded Reload
9892
WORD $0xd65f03c0 // ret
9993

10094
BB1_8:
@@ -114,25 +108,23 @@ BB1_9:
114108
WORD $0xf100116b // subs x11, x11, #4
115109
WORD $0x54ffff01 // b.ne LBB1_9
116110
WORD $0xeb02011f // cmp x8, x2
117-
WORD $0x54fffe00 // b.eq LBB1_7
118-
WORD $0x17ffffe5 // b LBB1_5
111+
WORD $0x54fffe20 // b.eq LBB1_7
112+
WORD $0x17ffffe6 // b LBB1_5
119113

120114
TEXT ·_or(SB), $0-32
121115
MOVD a+0(FP), R0
122116
MOVD b+8(FP), R1
123117
MOVD n+16(FP), R2
124-
WORD $0xa9bf7bfd // stp x29, x30, [sp, #-16]! ; 16-byte Folded Spill
125-
WORD $0x910003fd // mov x29, sp
126118
WORD $0xb40002a2 // cbz x2, LBB2_7
127119
WORD $0xf100105f // cmp x2, #4
128120
WORD $0x54000103 // b.lo LBB2_4
129121
WORD $0xd37df048 // lsl x8, x2, #3
130122
WORD $0x8b080029 // add x9, x1, x8
131123
WORD $0xeb00013f // cmp x9, x0
132-
WORD $0x54000229 // b.ls LBB2_8
124+
WORD $0x54000209 // b.ls LBB2_8
133125
WORD $0x8b080008 // add x8, x0, x8
134126
WORD $0xeb01011f // cmp x8, x1
135-
WORD $0x540001c9 // b.ls LBB2_8
127+
WORD $0x540001a9 // b.ls LBB2_8
136128

137129
BB2_4:
138130
WORD $0xd2800008 // mov x8, #0
@@ -152,7 +144,6 @@ BB2_6:
152144
WORD $0x54ffff61 // b.ne LBB2_6
153145

154146
BB2_7:
155-
WORD $0xa8c17bfd // ldp x29, x30, [sp], #16 ; 16-byte Folded Reload
156147
WORD $0xd65f03c0 // ret
157148

158149
BB2_8:
@@ -172,25 +163,23 @@ BB2_9:
172163
WORD $0xf100116b // subs x11, x11, #4
173164
WORD $0x54ffff01 // b.ne LBB2_9
174165
WORD $0xeb02011f // cmp x8, x2
175-
WORD $0x54fffe00 // b.eq LBB2_7
176-
WORD $0x17ffffe5 // b LBB2_5
166+
WORD $0x54fffe20 // b.eq LBB2_7
167+
WORD $0x17ffffe6 // b LBB2_5
177168

178169
TEXT ·_xor(SB), $0-32
179170
MOVD a+0(FP), R0
180171
MOVD b+8(FP), R1
181172
MOVD n+16(FP), R2
182-
WORD $0xa9bf7bfd // stp x29, x30, [sp, #-16]! ; 16-byte Folded Spill
183-
WORD $0x910003fd // mov x29, sp
184173
WORD $0xb40002a2 // cbz x2, LBB3_7
185174
WORD $0xf100105f // cmp x2, #4
186175
WORD $0x54000103 // b.lo LBB3_4
187176
WORD $0xd37df048 // lsl x8, x2, #3
188177
WORD $0x8b080029 // add x9, x1, x8
189178
WORD $0xeb00013f // cmp x9, x0
190-
WORD $0x54000229 // b.ls LBB3_8
179+
WORD $0x54000209 // b.ls LBB3_8
191180
WORD $0x8b080008 // add x8, x0, x8
192181
WORD $0xeb01011f // cmp x8, x1
193-
WORD $0x540001c9 // b.ls LBB3_8
182+
WORD $0x540001a9 // b.ls LBB3_8
194183

195184
BB3_4:
196185
WORD $0xd2800008 // mov x8, #0
@@ -210,7 +199,6 @@ BB3_6:
210199
WORD $0x54ffff61 // b.ne LBB3_6
211200

212201
BB3_7:
213-
WORD $0xa8c17bfd // ldp x29, x30, [sp], #16 ; 16-byte Folded Reload
214202
WORD $0xd65f03c0 // ret
215203

216204
BB3_8:
@@ -230,19 +218,17 @@ BB3_9:
230218
WORD $0xf100116b // subs x11, x11, #4
231219
WORD $0x54ffff01 // b.ne LBB3_9
232220
WORD $0xeb02011f // cmp x8, x2
233-
WORD $0x54fffe00 // b.eq LBB3_7
234-
WORD $0x17ffffe5 // b LBB3_5
221+
WORD $0x54fffe20 // b.eq LBB3_7
222+
WORD $0x17ffffe6 // b LBB3_5
235223

236224
TEXT ·_and_many(SB), $0-32
237225
MOVD a+0(FP), R0
238226
MOVD b+8(FP), R1
239227
MOVD dims+16(FP), R2
240-
WORD $0xa9bb67fa // stp x26, x25, [sp, #-80]! ; 16-byte Folded Spill
228+
WORD $0xa9bc67fa // stp x26, x25, [sp, #-64]! ; 16-byte Folded Spill
241229
WORD $0xa9015ff8 // stp x24, x23, [sp, #16] ; 16-byte Folded Spill
242230
WORD $0xa90257f6 // stp x22, x21, [sp, #32] ; 16-byte Folded Spill
243231
WORD $0xa9034ff4 // stp x20, x19, [sp, #48] ; 16-byte Folded Spill
244-
WORD $0xa9047bfd // stp x29, x30, [sp, #64] ; 16-byte Folded Spill
245-
WORD $0x910103fd // add x29, sp, #64
246232
WORD $0xf2407c48 // ands x8, x2, #0xffffffff
247233
WORD $0x54000a20 // b.eq LBB4_14
248234
WORD $0xd360fc4b // lsr x11, x2, #32
@@ -341,23 +327,20 @@ BB4_13:
341327
WORD $0x17ffffd9 // b LBB4_6
342328

343329
BB4_14:
344-
WORD $0xa9447bfd // ldp x29, x30, [sp, #64] ; 16-byte Folded Reload
345330
WORD $0xa9434ff4 // ldp x20, x19, [sp, #48] ; 16-byte Folded Reload
346331
WORD $0xa94257f6 // ldp x22, x21, [sp, #32] ; 16-byte Folded Reload
347332
WORD $0xa9415ff8 // ldp x24, x23, [sp, #16] ; 16-byte Folded Reload
348-
WORD $0xa8c567fa // ldp x26, x25, [sp], #80 ; 16-byte Folded Reload
333+
WORD $0xa8c467fa // ldp x26, x25, [sp], #64 ; 16-byte Folded Reload
349334
WORD $0xd65f03c0 // ret
350335

351336
TEXT ·_andn_many(SB), $0-32
352337
MOVD a+0(FP), R0
353338
MOVD b+8(FP), R1
354339
MOVD dims+16(FP), R2
355-
WORD $0xa9bb67fa // stp x26, x25, [sp, #-80]! ; 16-byte Folded Spill
340+
WORD $0xa9bc67fa // stp x26, x25, [sp, #-64]! ; 16-byte Folded Spill
356341
WORD $0xa9015ff8 // stp x24, x23, [sp, #16] ; 16-byte Folded Spill
357342
WORD $0xa90257f6 // stp x22, x21, [sp, #32] ; 16-byte Folded Spill
358343
WORD $0xa9034ff4 // stp x20, x19, [sp, #48] ; 16-byte Folded Spill
359-
WORD $0xa9047bfd // stp x29, x30, [sp, #64] ; 16-byte Folded Spill
360-
WORD $0x910103fd // add x29, sp, #64
361344
WORD $0xf2407c48 // ands x8, x2, #0xffffffff
362345
WORD $0x54000a20 // b.eq LBB5_14
363346
WORD $0xd360fc4b // lsr x11, x2, #32
@@ -456,23 +439,20 @@ BB5_13:
456439
WORD $0x17ffffd9 // b LBB5_6
457440

458441
BB5_14:
459-
WORD $0xa9447bfd // ldp x29, x30, [sp, #64] ; 16-byte Folded Reload
460442
WORD $0xa9434ff4 // ldp x20, x19, [sp, #48] ; 16-byte Folded Reload
461443
WORD $0xa94257f6 // ldp x22, x21, [sp, #32] ; 16-byte Folded Reload
462444
WORD $0xa9415ff8 // ldp x24, x23, [sp, #16] ; 16-byte Folded Reload
463-
WORD $0xa8c567fa // ldp x26, x25, [sp], #80 ; 16-byte Folded Reload
445+
WORD $0xa8c467fa // ldp x26, x25, [sp], #64 ; 16-byte Folded Reload
464446
WORD $0xd65f03c0 // ret
465447

466448
TEXT ·_or_many(SB), $0-32
467449
MOVD a+0(FP), R0
468450
MOVD b+8(FP), R1
469451
MOVD dims+16(FP), R2
470-
WORD $0xa9bb67fa // stp x26, x25, [sp, #-80]! ; 16-byte Folded Spill
452+
WORD $0xa9bc67fa // stp x26, x25, [sp, #-64]! ; 16-byte Folded Spill
471453
WORD $0xa9015ff8 // stp x24, x23, [sp, #16] ; 16-byte Folded Spill
472454
WORD $0xa90257f6 // stp x22, x21, [sp, #32] ; 16-byte Folded Spill
473455
WORD $0xa9034ff4 // stp x20, x19, [sp, #48] ; 16-byte Folded Spill
474-
WORD $0xa9047bfd // stp x29, x30, [sp, #64] ; 16-byte Folded Spill
475-
WORD $0x910103fd // add x29, sp, #64
476456
WORD $0xf2407c48 // ands x8, x2, #0xffffffff
477457
WORD $0x54000a20 // b.eq LBB6_14
478458
WORD $0xd360fc4b // lsr x11, x2, #32
@@ -571,23 +551,20 @@ BB6_13:
571551
WORD $0x17ffffd9 // b LBB6_6
572552

573553
BB6_14:
574-
WORD $0xa9447bfd // ldp x29, x30, [sp, #64] ; 16-byte Folded Reload
575554
WORD $0xa9434ff4 // ldp x20, x19, [sp, #48] ; 16-byte Folded Reload
576555
WORD $0xa94257f6 // ldp x22, x21, [sp, #32] ; 16-byte Folded Reload
577556
WORD $0xa9415ff8 // ldp x24, x23, [sp, #16] ; 16-byte Folded Reload
578-
WORD $0xa8c567fa // ldp x26, x25, [sp], #80 ; 16-byte Folded Reload
557+
WORD $0xa8c467fa // ldp x26, x25, [sp], #64 ; 16-byte Folded Reload
579558
WORD $0xd65f03c0 // ret
580559

581560
TEXT ·_xor_many(SB), $0-32
582561
MOVD a+0(FP), R0
583562
MOVD b+8(FP), R1
584563
MOVD dims+16(FP), R2
585-
WORD $0xa9bb67fa // stp x26, x25, [sp, #-80]! ; 16-byte Folded Spill
564+
WORD $0xa9bc67fa // stp x26, x25, [sp, #-64]! ; 16-byte Folded Spill
586565
WORD $0xa9015ff8 // stp x24, x23, [sp, #16] ; 16-byte Folded Spill
587566
WORD $0xa90257f6 // stp x22, x21, [sp, #32] ; 16-byte Folded Spill
588567
WORD $0xa9034ff4 // stp x20, x19, [sp, #48] ; 16-byte Folded Spill
589-
WORD $0xa9047bfd // stp x29, x30, [sp, #64] ; 16-byte Folded Spill
590-
WORD $0x910103fd // add x29, sp, #64
591568
WORD $0xf2407c48 // ands x8, x2, #0xffffffff
592569
WORD $0x54000a20 // b.eq LBB7_14
593570
WORD $0xd360fc4b // lsr x11, x2, #32
@@ -686,19 +663,16 @@ BB7_13:
686663
WORD $0x17ffffd9 // b LBB7_6
687664

688665
BB7_14:
689-
WORD $0xa9447bfd // ldp x29, x30, [sp, #64] ; 16-byte Folded Reload
690666
WORD $0xa9434ff4 // ldp x20, x19, [sp, #48] ; 16-byte Folded Reload
691667
WORD $0xa94257f6 // ldp x22, x21, [sp, #32] ; 16-byte Folded Reload
692668
WORD $0xa9415ff8 // ldp x24, x23, [sp, #16] ; 16-byte Folded Reload
693-
WORD $0xa8c567fa // ldp x26, x25, [sp], #80 ; 16-byte Folded Reload
669+
WORD $0xa8c467fa // ldp x26, x25, [sp], #64 ; 16-byte Folded Reload
694670
WORD $0xd65f03c0 // ret
695671

696672
TEXT ·_count(SB), $0-32
697673
MOVD a+0(FP), R0
698674
MOVD size+8(FP), R1
699675
MOVD result+16(FP), R2
700-
WORD $0xa9bf7bfd // stp x29, x30, [sp, #-16]! ; 16-byte Folded Spill
701-
WORD $0x910003fd // mov x29, sp
702676
WORD $0xb40000c1 // cbz x1, LBB8_3
703677
WORD $0xf100103f // cmp x1, #4
704678
WORD $0x540000c2 // b.hs LBB8_4
@@ -707,8 +681,8 @@ TEXT ·_count(SB), $0-32
707681
WORD $0x14000019 // b LBB8_7
708682

709683
BB8_3:
710-
WORD $0xd2800009 // mov x9, #0
711-
WORD $0x14000020 // b LBB8_9
684+
WORD $0xf900005f // str xzr, [x2]
685+
WORD $0xd65f03c0 // ret
712686

713687
BB8_4:
714688
WORD $0x927ef428 // and x8, x1, #0xfffffffffffffffc
@@ -751,5 +725,4 @@ BB8_8:
751725

752726
BB8_9:
753727
WORD $0xf9000049 // str x9, [x2]
754-
WORD $0xa8c17bfd // ldp x29, x30, [sp], #16 ; 16-byte Folded Reload
755728
WORD $0xd65f03c0 // ret

0 commit comments

Comments
 (0)