atomic_maybe_uninit/arch/
x86.rs

1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197
198
199
200
201
202
203
204
205
206
207
208
209
210
211
212
213
214
215
216
217
218
219
220
221
222
223
224
225
226
227
228
229
230
231
232
233
234
235
236
237
238
239
240
241
242
243
244
245
246
247
248
249
250
251
252
253
254
255
256
257
258
259
260
261
262
263
264
265
266
267
268
269
270
271
272
273
274
275
276
277
278
279
280
281
282
283
284
285
286
287
288
289
290
291
292
293
294
295
296
297
298
299
300
301
302
303
304
305
306
307
308
309
310
311
312
313
314
315
316
317
318
319
320
321
322
323
324
325
326
327
328
329
330
331
332
333
334
335
336
337
338
339
340
341
342
343
344
345
346
347
348
349
350
351
352
353
354
355
356
357
358
359
360
361
362
363
364
365
366
367
368
369
370
371
372
373
374
375
376
377
378
379
380
381
382
383
384
385
386
387
388
389
390
391
392
393
394
395
396
397
398
399
400
401
402
403
404
405
406
407
408
409
410
411
412
413
414
415
416
417
418
419
420
421
422
423
424
425
426
427
428
429
430
431
432
433
434
435
436
437
438
439
440
441
442
443
444
445
446
447
448
449
450
451
452
453
454
455
456
457
458
459
460
461
462
463
464
465
466
467
468
469
470
471
472
473
474
475
476
477
478
479
480
481
482
483
484
485
486
487
488
489
490
491
492
493
494
495
496
497
498
499
500
501
502
503
504
505
506
507
508
509
510
511
512
513
514
515
516
517
518
519
520
521
522
523
524
525
526
527
528
529
530
531
532
533
534
535
536
537
538
539
540
541
542
543
544
545
546
547
548
549
550
551
552
553
554
555
556
557
558
559
560
561
562
563
564
565
566
567
568
569
570
571
572
573
574
575
576
577
578
579
580
581
582
583
584
585
586
587
588
589
590
591
592
593
594
595
596
597
598
599
600
601
602
603
604
605
606
607
608
609
610
611
612
613
614
615
616
617
618
619
620
621
622
623
624
625
626
627
628
629
630
631
632
633
634
635
636
637
638
639
640
641
642
643
644
645
646
647
648
649
650
651
652
653
654
655
656
657
658
659
660
661
662
663
664
665
666
667
668
669
670
671
672
673
674
675
676
677
678
679
680
681
682
683
684
685
686
687
688
// SPDX-License-Identifier: Apache-2.0 OR MIT

/*
x86 and x86_64

Refs:
- x86 and amd64 instruction reference https://www.felixcloutier.com/x86
- portable-atomic https://github.com/taiki-e/portable-atomic

Generated asm:
- x86_64 https://godbolt.org/z/xKzj4WcaE
- x86_64 (+cmpxchg16b) https://godbolt.org/z/jzMoM9nhq
- x86 (i686) https://godbolt.org/z/sM6MPjYWf
- x86 (i686,-sse2) https://godbolt.org/z/MsrxfbcMG
- x86 (i586) https://godbolt.org/z/KEo6P7YEo
- x86 (i586,-x87) https://godbolt.org/z/P8cdjY7h1
*/

#[path = "cfgs/x86.rs"]
mod cfgs;

use core::{
    arch::asm,
    mem::{self, MaybeUninit},
    sync::atomic::Ordering,
};

use crate::raw::{AtomicCompareExchange, AtomicLoad, AtomicStore, AtomicSwap};
#[cfg(target_arch = "x86_64")]
#[cfg(any(target_feature = "cmpxchg16b", atomic_maybe_uninit_target_feature = "cmpxchg16b"))]
use crate::utils::{MaybeUninit128, Pair};
#[cfg(target_arch = "x86")]
#[cfg(not(atomic_maybe_uninit_no_cmpxchg8b))]
use crate::utils::{MaybeUninit64, Pair};

#[cfg(target_pointer_width = "32")]
macro_rules! ptr_modifier {
    () => {
        ":e"
    };
}
#[cfg(target_pointer_width = "64")]
macro_rules! ptr_modifier {
    () => {
        ""
    };
}

macro_rules! atomic {
    (
        $ty:ident, $val_reg:tt, $val_modifier:tt, $ptr_size:tt, $cmpxchg_cmp_reg:tt,
        $new_reg:tt
    ) => {
        impl AtomicLoad for $ty {
            #[inline]
            unsafe fn atomic_load(
                src: *const MaybeUninit<Self>,
                _order: Ordering,
            ) -> MaybeUninit<Self> {
                debug_assert!(src as usize % mem::size_of::<$ty>() == 0);
                let out: MaybeUninit<Self>;

                // SAFETY: the caller must uphold the safety contract.
                unsafe {
                    // atomic load is always SeqCst.
                    asm!(
                        concat!("mov {out", $val_modifier, "}, ", $ptr_size, " ptr [{src", ptr_modifier!(), "}]"), // atomic { out = *src }
                        src = in(reg) src,
                        out = lateout($val_reg) out,
                        options(nostack, preserves_flags),
                    );
                }
                out
            }
        }
        impl AtomicStore for $ty {
            #[inline]
            unsafe fn atomic_store(
                dst: *mut MaybeUninit<Self>,
                val: MaybeUninit<Self>,
                order: Ordering,
            ) {
                debug_assert!(dst as usize % mem::size_of::<$ty>() == 0);

                // SAFETY: the caller must uphold the safety contract.
                unsafe {
                    match order {
                        // Relaxed and Release stores are equivalent.
                        Ordering::Relaxed | Ordering::Release => {
                            asm!(
                                concat!("mov ", $ptr_size, " ptr [{dst", ptr_modifier!(), "}], {val", $val_modifier, "}"), // atomic { *dst = val }
                                dst = in(reg) dst,
                                val = in($val_reg) val,
                                options(nostack, preserves_flags),
                            );
                        }
                        Ordering::SeqCst => {
                            asm!(
                                // SeqCst store is xchg, not mov
                                concat!("xchg ", $ptr_size, " ptr [{dst", ptr_modifier!(), "}], {val", $val_modifier, "}"), // atomic { _x = *dst; *dst = val; val = _x }
                                dst = in(reg) dst,
                                val = inout($val_reg) val => _,
                                options(nostack, preserves_flags),
                            );
                        }
                        _ => unreachable!(),
                    }
                }
            }
        }
        impl AtomicSwap for $ty {
            #[inline]
            unsafe fn atomic_swap(
                dst: *mut MaybeUninit<Self>,
                val: MaybeUninit<Self>,
                _order: Ordering,
            ) -> MaybeUninit<Self> {
                debug_assert!(dst as usize % mem::size_of::<$ty>() == 0);
                let out: MaybeUninit<Self>;

                // SAFETY: the caller must uphold the safety contract.
                unsafe {
                    // atomic swap is always SeqCst.
                    asm!(
                        concat!("xchg ", $ptr_size, " ptr [{dst", ptr_modifier!(), "}], {val", $val_modifier, "}"), // atomic { _x = *dst; *dst = val; val = _x }
                        dst = in(reg) dst,
                        val = inout($val_reg) val => out,
                        options(nostack, preserves_flags),
                    );
                }
                out
            }
        }
        #[cfg(not(all(target_arch = "x86", atomic_maybe_uninit_no_cmpxchg)))]
        impl AtomicCompareExchange for $ty {
            #[inline]
            unsafe fn atomic_compare_exchange(
                dst: *mut MaybeUninit<Self>,
                old: MaybeUninit<Self>,
                new: MaybeUninit<Self>,
                _success: Ordering,
                _failure: Ordering,
            ) -> (MaybeUninit<Self>, bool) {
                debug_assert!(dst as usize % mem::size_of::<$ty>() == 0);
                let out: MaybeUninit<Self>;

                // SAFETY: the caller must uphold the safety contract.
                //
                // Refs: https://www.felixcloutier.com/x86/cmpxchg
                unsafe {
                    let r: u8;
                    // compare_exchange is always SeqCst.
                    asm!(
                        concat!("lock cmpxchg ", $ptr_size, " ptr [{dst", ptr_modifier!(), "}], ", $new_reg), // atomic { if *dst == $cmpxchg_cmp_reg { ZF = 1; *dst = $new_reg } else { ZF = 0; $cmpxchg_cmp_reg = *dst } }
                        "sete cl",                                                                            // cl = ZF
                        dst = in(reg) dst,
                        in($new_reg) new,
                        lateout("cl") r,
                        inout($cmpxchg_cmp_reg) old => out,
                        // Do not use `preserves_flags` because CMPXCHG modifies the ZF, CF, PF, AF, SF, and OF flags.
                        options(nostack),
                    );
                    crate::utils::assert_unchecked(r == 0 || r == 1); // may help remove extra test
                    (out, r != 0)
                }
            }
        }
    };
}

atomic!(i8, reg_byte, "", "byte", "al", "cl");
atomic!(u8, reg_byte, "", "byte", "al", "cl");
atomic!(i16, reg, ":x", "word", "ax", "cx");
atomic!(u16, reg, ":x", "word", "ax", "cx");
atomic!(i32, reg, ":e", "dword", "eax", "ecx");
atomic!(u32, reg, ":e", "dword", "eax", "ecx");
#[cfg(target_arch = "x86_64")]
atomic!(i64, reg, "", "qword", "rax", "rcx");
#[cfg(target_arch = "x86_64")]
atomic!(u64, reg, "", "qword", "rax", "rcx");
#[cfg(target_pointer_width = "32")]
atomic!(isize, reg, ":e", "dword", "eax", "ecx");
#[cfg(target_pointer_width = "32")]
atomic!(usize, reg, ":e", "dword", "eax", "ecx");
#[cfg(target_pointer_width = "64")]
atomic!(isize, reg, "", "qword", "rax", "rcx");
#[cfg(target_pointer_width = "64")]
atomic!(usize, reg, "", "qword", "rax", "rcx");

// For load/store, we can use MOVQ(SSE2)/MOVLPS(SSE)/FILD&FISTP(x87) instead of CMPXCHG8B.
// Refs: https://github.com/llvm/llvm-project/blob/llvmorg-19.1.0/llvm/test/CodeGen/X86/atomic-load-store-wide.ll
#[cfg(target_arch = "x86")]
#[cfg(not(atomic_maybe_uninit_no_cmpxchg8b))]
macro_rules! atomic64 {
    ($ty:ident) => {
        impl AtomicLoad for $ty {
            #[inline]
            unsafe fn atomic_load(
                src: *const MaybeUninit<Self>,
                _order: Ordering,
            ) -> MaybeUninit<Self> {
                debug_assert!(src as usize % mem::size_of::<$ty>() == 0);

                #[cfg(target_feature = "sse2")]
                // SAFETY: the caller must uphold the safety contract.
                // cfg guarantees that the CPU supports SSE.
                //
                // Refs:
                // - https://www.felixcloutier.com/x86/movq (SSE2)
                // - https://www.felixcloutier.com/x86/movd:movq (SSE2)
                unsafe {
                    let out: MaybeUninit<core::arch::x86::__m128i>;
                    // atomic load is always SeqCst.
                    asm!(
                        "movq {out}, qword ptr [{src}]", // atomic { out[:] = *src }
                        src = in(reg) src,
                        out = out(xmm_reg) out,
                        options(nostack, preserves_flags),
                    );
                    #[allow(clippy::missing_transmute_annotations)] // false positive: out is already type annotated
                    core::mem::transmute::<_, [MaybeUninit<Self>; 2]>(out)[0]
                }
                #[cfg(not(target_feature = "sse2"))]
                #[cfg(target_feature = "sse")]
                // SAFETY: the caller must uphold the safety contract.
                // cfg guarantees that the CPU supports SSE.
                //
                // Refs:
                // - https://www.felixcloutier.com/x86/movlps (SSE)
                unsafe {
                    let mut out = MaybeUninit::<core::arch::x86::__m128>::zeroed();
                    // atomic load is always SeqCst.
                    asm!(
                        "movlps {out}, qword ptr [{src}]", // atomic { out[:] = *src }
                        src = in(reg) src,
                        out = inout(xmm_reg) out,
                        options(nostack, preserves_flags),
                    );
                    #[allow(clippy::missing_transmute_annotations)] // false positive: out is already type annotated
                    core::mem::transmute::<_, [MaybeUninit<Self>; 2]>(out)[0]
                }
                #[cfg(not(target_feature = "sse"))]
                #[cfg(any(target_feature = "x87", atomic_maybe_uninit_target_feature = "x87"))]
                // SAFETY: the caller must uphold the safety contract.
                //
                // Refs:
                // - https://www.felixcloutier.com/x86/fild
                // - https://www.felixcloutier.com/x86/fist:fistp
                unsafe {
                    let mut out = MaybeUninit::<Self>::uninit();
                    // atomic load is always SeqCst.
                    asm!(
                        "fild qword ptr [{src}]",  // atomic { st.push(*src) }
                        "fistp qword ptr [{out}]", // *out = st.pop()
                        src = in(reg) src,
                        out = in(reg) out.as_mut_ptr(),
                        out("st(0)") _,
                        out("st(1)") _,
                        out("st(2)") _,
                        out("st(3)") _,
                        out("st(4)") _,
                        out("st(5)") _,
                        out("st(6)") _,
                        out("st(7)") _,
                        // Do not use `preserves_flags` because FILD and FISTP modify C1 in x87 FPU status word.
                        options(nostack),
                    );
                    out
                }
                #[cfg(not(target_feature = "sse"))]
                #[cfg(not(any(target_feature = "x87", atomic_maybe_uninit_target_feature = "x87")))]
                // SAFETY: the caller must uphold the safety contract.
                //
                // Refs: https://www.felixcloutier.com/x86/cmpxchg8b:cmpxchg16b
                unsafe {
                    let (prev_lo, prev_hi);
                    // atomic load is always SeqCst.
                    asm!(
                        "lock cmpxchg8b qword ptr [edi]", // atomic { if *edi == edx:eax { ZF = 1; *edi = ecx:ebx } else { ZF = 0; edx:eax = *edi } }
                        // set old/new args of CMPXCHG8B to 0
                        in("ebx") 0_u32,
                        in("ecx") 0_u32,
                        inout("eax") 0_u32 => prev_lo,
                        inout("edx") 0_u32 => prev_hi,
                        in("edi") src,
                        // Do not use `preserves_flags` because CMPXCHG8B modifies the ZF flag.
                        options(nostack),
                    );
                    MaybeUninit64 { pair: Pair { lo: prev_lo, hi: prev_hi } }.$ty
                }
            }
        }
        impl AtomicStore for $ty {
            #[inline]
            unsafe fn atomic_store(
                dst: *mut MaybeUninit<Self>,
                val: MaybeUninit<Self>,
                order: Ordering,
            ) {
                debug_assert!(dst as usize % mem::size_of::<$ty>() == 0);

                #[cfg(target_feature = "sse")]
                // SAFETY: the caller must uphold the safety contract.
                // cfg guarantees that the CPU supports SSE.
                //
                // Refs:
                // - https://www.felixcloutier.com/x86/movlps (SSE)
                // - https://www.felixcloutier.com/x86/lock
                // - https://www.felixcloutier.com/x86/or
                unsafe {
                    let val: MaybeUninit<core::arch::x86::__m128>
                        = core::mem::transmute([val, MaybeUninit::uninit()]);
                    match order {
                        // Relaxed and Release stores are equivalent.
                        Ordering::Relaxed | Ordering::Release => {
                            asm!(
                                "movlps qword ptr [{dst}], {val}", // atomic { *dst = val[:] }
                                dst = in(reg) dst,
                                val = in(xmm_reg) val,
                                options(nostack, preserves_flags),
                            );
                        }
                        Ordering::SeqCst => {
                            asm!(
                                "movlps qword ptr [{dst}], {val}", // atomic { *dst = val[:] }
                                // equivalent to mfence, but doesn't require SSE2
                                "lock or dword ptr [esp], 0",      // fence
                                dst = in(reg) dst,
                                val = in(xmm_reg) val,
                                // Do not use `preserves_flags` because OR modifies the OF, CF, SF, ZF, and PF flags.
                                options(nostack),
                            );
                        }
                        _ => unreachable!(),
                    }
                }
                #[cfg(not(target_feature = "sse"))]
                #[cfg(any(target_feature = "x87", atomic_maybe_uninit_target_feature = "x87"))]
                // SAFETY: the caller must uphold the safety contract.
                //
                // Refs:
                // - https://www.felixcloutier.com/x86/fild
                // - https://www.felixcloutier.com/x86/fist:fistp
                unsafe {
                    match order {
                        // Relaxed and Release stores are equivalent.
                        Ordering::Relaxed | Ordering::Release => {
                            asm!(
                                "fild qword ptr [{val}]",  // st.push(*val)
                                "fistp qword ptr [{dst}]", // atomic { *dst = st.pop() }
                                val = in(reg) val.as_ptr(),
                                dst = in(reg) dst,
                                out("st(0)") _,
                                out("st(1)") _,
                                out("st(2)") _,
                                out("st(3)") _,
                                out("st(4)") _,
                                out("st(5)") _,
                                out("st(6)") _,
                                out("st(7)") _,
                                // Do not use `preserves_flags` because FILD and FISTP modify condition code flags in x87 FPU status word.
                                options(nostack),
                            );
                        }
                        Ordering::SeqCst => {
                            asm!(
                                "fild qword ptr [{val}]",     // st.push(*val)
                                "fistp qword ptr [{dst}]",    // atomic { *dst = st.pop() }
                                // equivalent to mfence, but doesn't require SSE2
                                "lock or dword ptr [esp], 0", // fence
                                val = in(reg) val.as_ptr(),
                                dst = in(reg) dst,
                                out("st(0)") _,
                                out("st(1)") _,
                                out("st(2)") _,
                                out("st(3)") _,
                                out("st(4)") _,
                                out("st(5)") _,
                                out("st(6)") _,
                                out("st(7)") _,
                                // Do not use `preserves_flags` because OR modifies the OF, CF, SF, ZF, and PF flags, FILD and FISTP modify condition code flags in x87 FPU status word.
                                options(nostack),
                            );
                        }
                        _ => unreachable!(),
                    }
                }
                #[cfg(not(target_feature = "sse"))]
                #[cfg(not(any(target_feature = "x87", atomic_maybe_uninit_target_feature = "x87")))]
                // SAFETY: the caller must uphold the safety contract.
                //
                // Refs: https://www.felixcloutier.com/x86/cmpxchg8b:cmpxchg16b
                unsafe {
                    let val = MaybeUninit64 { $ty: val };
                    // atomic store by CMPXCHG8B is always SeqCst.
                    let _ = order;
                    asm!(
                        // This is based on the code generated for the first load in DW RMWs by LLVM,
                        // but it is interesting that they generate code that does mixed-sized atomic access.
                        //
                        // This is not single-copy atomic reads, but this is ok because subsequent
                        // CAS will check for consistency.
                        "mov eax, dword ptr [edi]",           // atomic { eax = *edi }
                        "mov edx, dword ptr [edi + 4]",       // atomic { edx = *edi.add(4) }
                        "2:", // 'retry:
                            "lock cmpxchg8b qword ptr [edi]", // atomic { if *edi == edx:eax { ZF = 1; *edi = ecx:ebx } else { ZF = 0; edx:eax = *edi } }
                            "jne 2b",                         // if ZF == 0 { jump 'retry }
                        in("ebx") val.pair.lo,
                        in("ecx") val.pair.hi,
                        out("eax") _,
                        out("edx") _,
                        in("edi") dst,
                        // Do not use `preserves_flags` because CMPXCHG8B modifies the ZF flag.
                        options(nostack),
                    );
                }
            }
        }
        impl AtomicSwap for $ty {
            #[inline]
            unsafe fn atomic_swap(
                dst: *mut MaybeUninit<Self>,
                val: MaybeUninit<Self>,
                _order: Ordering,
            ) -> MaybeUninit<Self> {
                debug_assert!(dst as usize % mem::size_of::<$ty>() == 0);
                let val = MaybeUninit64 { $ty: val };
                let (mut prev_lo, mut prev_hi);

                // SAFETY: the caller must uphold the safety contract.
                //
                // Refs: https://www.felixcloutier.com/x86/cmpxchg8b:cmpxchg16b
                unsafe {
                    // atomic swap is always SeqCst.
                    asm!(
                        // This is based on the code generated for the first load in DW RMWs by LLVM,
                        // but it is interesting that they generate code that does mixed-sized atomic access.
                        //
                        // This is not single-copy atomic reads, but this is ok because subsequent
                        // CAS will check for consistency.
                        "mov eax, dword ptr [edi]",           // atomic { eax = *edi }
                        "mov edx, dword ptr [edi + 4]",       // atomic { edx = *edi.add(4) }
                        "2:", // 'retry:
                            "lock cmpxchg8b qword ptr [edi]", // atomic { if *edi == edx:eax { ZF = 1; *edi = ecx:ebx } else { ZF = 0; edx:eax = *edi } }
                            "jne 2b",                         // if ZF == 0 { jump 'retry }
                        in("ebx") val.pair.lo,
                        in("ecx") val.pair.hi,
                        out("eax") prev_lo,
                        out("edx") prev_hi,
                        in("edi") dst,
                        // Do not use `preserves_flags` because CMPXCHG8B modifies the ZF flag.
                        options(nostack),
                    );
                    MaybeUninit64 { pair: Pair { lo: prev_lo, hi: prev_hi } }.$ty
                }
            }
        }
        impl AtomicCompareExchange for $ty {
            #[inline]
            unsafe fn atomic_compare_exchange(
                dst: *mut MaybeUninit<Self>,
                old: MaybeUninit<Self>,
                new: MaybeUninit<Self>,
                _success: Ordering,
                _failure: Ordering,
            ) -> (MaybeUninit<Self>, bool) {
                debug_assert!(dst as usize % mem::size_of::<$ty>() == 0);
                let old = MaybeUninit64 { $ty: old };
                let new = MaybeUninit64 { $ty: new };
                let (prev_lo, prev_hi);

                // SAFETY: the caller must uphold the safety contract.
                //
                // Refs: https://www.felixcloutier.com/x86/cmpxchg8b:cmpxchg16b
                unsafe {
                    let r: u8;
                    // compare_exchange is always SeqCst.
                    asm!(
                        "lock cmpxchg8b qword ptr [edi]", // atomic { if *edi == edx:eax { ZF = 1; *edi = ecx:ebx } else { ZF = 0; edx:eax = *edi } }
                        "sete cl",                        // cl = ZF
                        in("ebx") new.pair.lo,
                        in("ecx") new.pair.hi,
                        inout("eax") old.pair.lo => prev_lo,
                        inout("edx") old.pair.hi => prev_hi,
                        in("edi") dst,
                        lateout("cl") r,
                        // Do not use `preserves_flags` because CMPXCHG8B modifies the ZF flag.
                        options(nostack),
                    );
                    crate::utils::assert_unchecked(r == 0 || r == 1); // may help remove extra test
                    (
                        MaybeUninit64 { pair: Pair { lo: prev_lo, hi: prev_hi } }.$ty,
                        r != 0
                    )
                }
            }
        }
    };
}

#[cfg(target_arch = "x86")]
#[cfg(not(atomic_maybe_uninit_no_cmpxchg8b))]
atomic64!(i64);
#[cfg(target_arch = "x86")]
#[cfg(not(atomic_maybe_uninit_no_cmpxchg8b))]
atomic64!(u64);

#[cfg(target_arch = "x86_64")]
#[cfg(any(target_feature = "cmpxchg16b", atomic_maybe_uninit_target_feature = "cmpxchg16b"))]
macro_rules! atomic128 {
    ($ty:ident) => {
        #[cfg(target_pointer_width = "32")]
        atomic128!($ty, "edi");
        #[cfg(target_pointer_width = "64")]
        atomic128!($ty, "rdi");
    };
    ($ty:ident, $rdi:tt) => {
        impl AtomicLoad for $ty {
            #[inline]
            unsafe fn atomic_load(
                src: *const MaybeUninit<Self>,
                _order: Ordering,
            ) -> MaybeUninit<Self> {
                debug_assert!(src as usize % mem::size_of::<$ty>() == 0);
                let (prev_lo, prev_hi);

                // SAFETY: the caller must guarantee that `src` is valid for both writes and
                // reads, 16-byte aligned, and that there are no concurrent non-atomic operations.
                // cfg guarantees that the CPU supports CMPXCHG16B.
                //
                // Refs: https://www.felixcloutier.com/x86/cmpxchg8b:cmpxchg16b
                unsafe {
                    // atomic load is always SeqCst.
                    asm!(
                        "mov {rbx_tmp}, rbx", // save rbx which is reserved by LLVM
                        "xor rbx, rbx",       // zeroed rbx
                        concat!("lock cmpxchg16b xmmword ptr [", $rdi, "]"), // atomic { if *$rdi == rdx:rax { ZF = 1; *$rdi = rcx:rbx } else { ZF = 0; rdx:rax = *$rdi } }
                        "mov rbx, {rbx_tmp}", // restore rbx
                        // set old/new args of CMPXCHG16B to 0 (rbx is zeroed after saved to rbx_tmp, to avoid xchg)
                        rbx_tmp = out(reg) _,
                        in("rcx") 0_u64,
                        inout("rax") 0_u64 => prev_lo,
                        inout("rdx") 0_u64 => prev_hi,
                        in($rdi) src,
                        // Do not use `preserves_flags` because CMPXCHG16B modifies the ZF flag.
                        options(nostack),
                    );
                    MaybeUninit128 { pair: Pair { lo: prev_lo, hi: prev_hi } }.$ty
                }
            }
        }
        impl AtomicStore for $ty {
            #[inline]
            unsafe fn atomic_store(
                dst: *mut MaybeUninit<Self>,
                val: MaybeUninit<Self>,
                _order: Ordering,
            ) {
                debug_assert!(dst as usize % mem::size_of::<$ty>() == 0);
                let val = MaybeUninit128 { $ty: val };

                // SAFETY: the caller must guarantee that `dst` is valid for both writes and
                // reads, 16-byte aligned, and that there are no concurrent non-atomic operations.
                // cfg guarantees that the CPU supports CMPXCHG16B.
                //
                // Refs: https://www.felixcloutier.com/x86/cmpxchg8b:cmpxchg16b
                unsafe {
                    // atomic store is always SeqCst.
                    asm!(
                        "xchg {rbx_tmp}, rbx", // save rbx which is reserved by LLVM
                        // This is based on the code generated for the first load in DW RMWs by LLVM,
                        // but it is interesting that they generate code that does mixed-sized atomic access.
                        //
                        // This is not single-copy atomic reads, but this is ok because subsequent
                        // CAS will check for consistency.
                        concat!("mov rax, qword ptr [", $rdi, "]"),              // atomic { rax = *$rdi }
                        concat!("mov rdx, qword ptr [", $rdi, " + 8]"),          // atomic { rdx = *$rdi.add(8) }
                        "2:", // 'retry:
                            concat!("lock cmpxchg16b xmmword ptr [", $rdi, "]"), // atomic { if *$rdi == rdx:rax { ZF = 1; *$rdi = rcx:rbx } else { ZF = 0; rdx:rax = *$rdi } }
                            "jne 2b",                                            // if ZF == 0 { jump 'retry }
                        "mov rbx, {rbx_tmp}", // restore rbx
                        rbx_tmp = inout(reg) val.pair.lo => _,
                        in("rcx") val.pair.hi,
                        out("rax") _,
                        out("rdx") _,
                        in($rdi) dst,
                        // Do not use `preserves_flags` because CMPXCHG16B modifies the ZF flag.
                        options(nostack),
                    );
                }
            }
        }
        impl AtomicSwap for $ty {
            #[inline]
            unsafe fn atomic_swap(
                dst: *mut MaybeUninit<Self>,
                val: MaybeUninit<Self>,
                _order: Ordering,
            ) -> MaybeUninit<Self> {
                debug_assert!(dst as usize % mem::size_of::<$ty>() == 0);
                let val = MaybeUninit128 { $ty: val };
                let (mut prev_lo, mut prev_hi);

                // SAFETY: the caller must guarantee that `dst` is valid for both writes and
                // reads, 16-byte aligned, and that there are no concurrent non-atomic operations.
                // cfg guarantees that the CPU supports CMPXCHG16B.
                //
                // Refs: https://www.felixcloutier.com/x86/cmpxchg8b:cmpxchg16b
                unsafe {
                    // atomic swap is always SeqCst.
                    asm!(
                        "xchg {rbx_tmp}, rbx", // save rbx which is reserved by LLVM
                        // This is based on the code generated for the first load in DW RMWs by LLVM,
                        // but it is interesting that they generate code that does mixed-sized atomic access.
                        //
                        // This is not single-copy atomic reads, but this is ok because subsequent
                        // CAS will check for consistency.
                        concat!("mov rax, qword ptr [", $rdi, "]"),              // atomic { rax = *$rdi }
                        concat!("mov rdx, qword ptr [", $rdi, " + 8]"),          // atomic { rdx = *$rdi.add(8) }
                        "2:", // 'retry:
                            concat!("lock cmpxchg16b xmmword ptr [", $rdi, "]"), // atomic { if *$rdi == rdx:rax { ZF = 1; *$rdi = rcx:rbx } else { ZF = 0; rdx:rax = *$rdi } }
                            "jne 2b",                                            // if ZF == 0 { jump 'retry }
                        "mov rbx, {rbx_tmp}", // restore rbx
                        rbx_tmp = inout(reg) val.pair.lo => _,
                        in("rcx") val.pair.hi,
                        out("rax") prev_lo,
                        out("rdx") prev_hi,
                        in($rdi) dst,
                        // Do not use `preserves_flags` because CMPXCHG16B modifies the ZF flag.
                        options(nostack),
                    );
                    MaybeUninit128 { pair: Pair { lo: prev_lo, hi: prev_hi } }.$ty
                }
            }
        }
        impl AtomicCompareExchange for $ty {
            #[inline]
            unsafe fn atomic_compare_exchange(
                dst: *mut MaybeUninit<Self>,
                old: MaybeUninit<Self>,
                new: MaybeUninit<Self>,
                _success: Ordering,
                _failure: Ordering,
            ) -> (MaybeUninit<Self>, bool) {
                debug_assert!(dst as usize % mem::size_of::<$ty>() == 0);
                let old = MaybeUninit128 { $ty: old };
                let new = MaybeUninit128 { $ty: new };
                let (prev_lo, prev_hi);

                // SAFETY: the caller must guarantee that `dst` is valid for both writes and
                // reads, 16-byte aligned, and that there are no concurrent non-atomic operations.
                // cfg guarantees that the CPU supports CMPXCHG16B.
                //
                // Refs: https://www.felixcloutier.com/x86/cmpxchg8b:cmpxchg16b
                unsafe {
                    let r: u8;
                    // compare_exchange is always SeqCst.
                    asm!(
                        "xchg {rbx_tmp}, rbx", // save rbx which is reserved by LLVM
                        concat!("lock cmpxchg16b xmmword ptr [", $rdi, "]"), // atomic { if *$rdi == rdx:rax { ZF = 1; *$rdi = rcx:rbx } else { ZF = 0; rdx:rax = *$rdi } }
                        "sete cl",                                           // cl = ZF
                        "mov rbx, {rbx_tmp}", // restore rbx
                        rbx_tmp = inout(reg) new.pair.lo => _,
                        in("rcx") new.pair.hi,
                        inout("rax") old.pair.lo => prev_lo,
                        inout("rdx") old.pair.hi => prev_hi,
                        in($rdi) dst,
                        lateout("cl") r,
                        // Do not use `preserves_flags` because CMPXCHG16B modifies the ZF flag.
                        options(nostack),
                    );
                    crate::utils::assert_unchecked(r == 0 || r == 1); // may help remove extra test
                    (
                        MaybeUninit128 { pair: Pair { lo: prev_lo, hi: prev_hi } }.$ty,
                        r != 0
                    )
                }
            }
        }
    };
}

#[cfg(target_arch = "x86_64")]
#[cfg(any(target_feature = "cmpxchg16b", atomic_maybe_uninit_target_feature = "cmpxchg16b"))]
atomic128!(i128);
#[cfg(target_arch = "x86_64")]
#[cfg(any(target_feature = "cmpxchg16b", atomic_maybe_uninit_target_feature = "cmpxchg16b"))]
atomic128!(u128);