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
689
690
691
692
693
694
695
696
697
698
699
700
701
702
703
704
705
706
707
708
709
710
711
712
713
714
715
716
717
718
719
720
721
722
723
724
725
726
727
728
729
730
731
732
733
734
735
736
737
738
739
740
741
742
743
744
745
746
747
748
749
750
751
752
753
754
755
756
757
758
759
760
761
762
763
764
765
766
767
768
769
770
771
772
773
774
775
776
777
778
779
780
781
782
783
784
785
786
787
788
789
790
791
792
793
794
795
796
797
798
799
800
801
802
803
804
805
806
807
808
809
810
811
812
813
814
815
816
817
818
819
820
821
822
823
824
825
826
827
828
829
830
831
832
833
834
835
836
837
838
839
840
841
842
// -*- mode: rust; -*-
//
// This file is part of curve25519-dalek.
// Copyright (c) 2016-2021 isis lovecruft
// Copyright (c) 2016-2019 Henry de Valence
// See LICENSE for licensing information.
//
// Authors:
// - isis agora lovecruft <isis@patternsinthevoid.net>
// - Henry de Valence <hdevalence@hdevalence.ca>

#![allow(non_snake_case)]

use crate::backend::vector::packed_simd::u64x4;
use core::ops::{Add, Mul, Neg};

use crate::backend::serial::u64::field::FieldElement51;

use curve25519_dalek_derive::unsafe_target_feature;

/// A wrapper around `vpmadd52luq` that works on `u64x4`.
#[unsafe_target_feature("avx512ifma,avx512vl")]
#[inline]
unsafe fn madd52lo(z: u64x4, x: u64x4, y: u64x4) -> u64x4 {
    use core::arch::x86_64::_mm256_madd52lo_epu64;
    _mm256_madd52lo_epu64(z.into(), x.into(), y.into()).into()
}

/// A wrapper around `vpmadd52huq` that works on `u64x4`.
#[unsafe_target_feature("avx512ifma,avx512vl")]
#[inline]
unsafe fn madd52hi(z: u64x4, x: u64x4, y: u64x4) -> u64x4 {
    use core::arch::x86_64::_mm256_madd52hi_epu64;
    _mm256_madd52hi_epu64(z.into(), x.into(), y.into()).into()
}

/// A vector of four field elements in radix 2^51, with unreduced coefficients.
#[derive(Copy, Clone, Debug)]
pub struct F51x4Unreduced(pub(crate) [u64x4; 5]);

/// A vector of four field elements in radix 2^51, with reduced coefficients.
#[derive(Copy, Clone, Debug)]
pub struct F51x4Reduced(pub(crate) [u64x4; 5]);

#[allow(clippy::upper_case_acronyms)]
#[derive(Copy, Clone)]
pub enum Shuffle {
    AAAA,
    BBBB,
    BADC,
    BACD,
    ADDA,
    CBCB,
    ABDC,
    ABAB,
    DBBD,
    CACA,
}

#[unsafe_target_feature("avx512ifma,avx512vl")]
#[inline(always)]
fn shuffle_lanes(x: u64x4, control: Shuffle) -> u64x4 {
    unsafe {
        use core::arch::x86_64::_mm256_permute4x64_epi64 as perm;

        match control {
            Shuffle::AAAA => perm(x.into(), 0b00_00_00_00).into(),
            Shuffle::BBBB => perm(x.into(), 0b01_01_01_01).into(),
            Shuffle::BADC => perm(x.into(), 0b10_11_00_01).into(),
            Shuffle::BACD => perm(x.into(), 0b11_10_00_01).into(),
            Shuffle::ADDA => perm(x.into(), 0b00_11_11_00).into(),
            Shuffle::CBCB => perm(x.into(), 0b01_10_01_10).into(),
            Shuffle::ABDC => perm(x.into(), 0b10_11_01_00).into(),
            Shuffle::ABAB => perm(x.into(), 0b01_00_01_00).into(),
            Shuffle::DBBD => perm(x.into(), 0b11_01_01_11).into(),
            Shuffle::CACA => perm(x.into(), 0b00_10_00_10).into(),
        }
    }
}

#[allow(clippy::upper_case_acronyms)]
#[derive(Copy, Clone)]
pub enum Lanes {
    D,
    C,
    AB,
    AC,
    AD,
    BCD,
}

#[unsafe_target_feature("avx512ifma,avx512vl")]
#[inline]
fn blend_lanes(x: u64x4, y: u64x4, control: Lanes) -> u64x4 {
    unsafe {
        use core::arch::x86_64::_mm256_blend_epi32 as blend;

        match control {
            Lanes::D => blend(x.into(), y.into(), 0b11_00_00_00).into(),
            Lanes::C => blend(x.into(), y.into(), 0b00_11_00_00).into(),
            Lanes::AB => blend(x.into(), y.into(), 0b00_00_11_11).into(),
            Lanes::AC => blend(x.into(), y.into(), 0b00_11_00_11).into(),
            Lanes::AD => blend(x.into(), y.into(), 0b11_00_00_11).into(),
            Lanes::BCD => blend(x.into(), y.into(), 0b11_11_11_00).into(),
        }
    }
}

#[unsafe_target_feature("avx512ifma,avx512vl")]
impl F51x4Unreduced {
    pub const ZERO: F51x4Unreduced = F51x4Unreduced([u64x4::splat_const::<0>(); 5]);

    pub fn new(
        x0: &FieldElement51,
        x1: &FieldElement51,
        x2: &FieldElement51,
        x3: &FieldElement51,
    ) -> F51x4Unreduced {
        F51x4Unreduced([
            u64x4::new(x0.0[0], x1.0[0], x2.0[0], x3.0[0]),
            u64x4::new(x0.0[1], x1.0[1], x2.0[1], x3.0[1]),
            u64x4::new(x0.0[2], x1.0[2], x2.0[2], x3.0[2]),
            u64x4::new(x0.0[3], x1.0[3], x2.0[3], x3.0[3]),
            u64x4::new(x0.0[4], x1.0[4], x2.0[4], x3.0[4]),
        ])
    }

    pub fn split(&self) -> [FieldElement51; 4] {
        let x = &self.0;
        [
            FieldElement51([
                x[0].extract::<0>(),
                x[1].extract::<0>(),
                x[2].extract::<0>(),
                x[3].extract::<0>(),
                x[4].extract::<0>(),
            ]),
            FieldElement51([
                x[0].extract::<1>(),
                x[1].extract::<1>(),
                x[2].extract::<1>(),
                x[3].extract::<1>(),
                x[4].extract::<1>(),
            ]),
            FieldElement51([
                x[0].extract::<2>(),
                x[1].extract::<2>(),
                x[2].extract::<2>(),
                x[3].extract::<2>(),
                x[4].extract::<2>(),
            ]),
            FieldElement51([
                x[0].extract::<3>(),
                x[1].extract::<3>(),
                x[2].extract::<3>(),
                x[3].extract::<3>(),
                x[4].extract::<3>(),
            ]),
        ]
    }

    #[inline]
    pub fn diff_sum(&self) -> F51x4Unreduced {
        // tmp1 = (B, A, D, C)
        let tmp1 = self.shuffle(Shuffle::BADC);
        // tmp2 = (-A, B, -C, D)
        let tmp2 = self.blend(&self.negate_lazy(), Lanes::AC);
        // (B - A, B + A, D - C, D + C)
        tmp1 + tmp2
    }

    #[inline]
    pub fn negate_lazy(&self) -> F51x4Unreduced {
        let lo = u64x4::splat(36028797018963664u64);
        let hi = u64x4::splat(36028797018963952u64);
        F51x4Unreduced([
            lo - self.0[0],
            hi - self.0[1],
            hi - self.0[2],
            hi - self.0[3],
            hi - self.0[4],
        ])
    }

    #[inline]
    pub fn shuffle(&self, control: Shuffle) -> F51x4Unreduced {
        F51x4Unreduced([
            shuffle_lanes(self.0[0], control),
            shuffle_lanes(self.0[1], control),
            shuffle_lanes(self.0[2], control),
            shuffle_lanes(self.0[3], control),
            shuffle_lanes(self.0[4], control),
        ])
    }

    #[inline]
    pub fn blend(&self, other: &F51x4Unreduced, control: Lanes) -> F51x4Unreduced {
        F51x4Unreduced([
            blend_lanes(self.0[0], other.0[0], control),
            blend_lanes(self.0[1], other.0[1], control),
            blend_lanes(self.0[2], other.0[2], control),
            blend_lanes(self.0[3], other.0[3], control),
            blend_lanes(self.0[4], other.0[4], control),
        ])
    }
}

#[unsafe_target_feature("avx512ifma,avx512vl")]
impl Neg for F51x4Reduced {
    type Output = F51x4Reduced;

    fn neg(self) -> F51x4Reduced {
        F51x4Unreduced::from(self).negate_lazy().into()
    }
}

use subtle::Choice;
use subtle::ConditionallySelectable;

#[unsafe_target_feature("avx512ifma,avx512vl")]
impl ConditionallySelectable for F51x4Reduced {
    #[inline]
    fn conditional_select(a: &F51x4Reduced, b: &F51x4Reduced, choice: Choice) -> F51x4Reduced {
        let mask = (-(choice.unwrap_u8() as i64)) as u64;
        let mask_vec = u64x4::splat(mask);
        F51x4Reduced([
            a.0[0] ^ (mask_vec & (a.0[0] ^ b.0[0])),
            a.0[1] ^ (mask_vec & (a.0[1] ^ b.0[1])),
            a.0[2] ^ (mask_vec & (a.0[2] ^ b.0[2])),
            a.0[3] ^ (mask_vec & (a.0[3] ^ b.0[3])),
            a.0[4] ^ (mask_vec & (a.0[4] ^ b.0[4])),
        ])
    }

    #[inline]
    fn conditional_assign(&mut self, other: &F51x4Reduced, choice: Choice) {
        let mask = (-(choice.unwrap_u8() as i64)) as u64;
        let mask_vec = u64x4::splat(mask);
        self.0[0] ^= mask_vec & (self.0[0] ^ other.0[0]);
        self.0[1] ^= mask_vec & (self.0[1] ^ other.0[1]);
        self.0[2] ^= mask_vec & (self.0[2] ^ other.0[2]);
        self.0[3] ^= mask_vec & (self.0[3] ^ other.0[3]);
        self.0[4] ^= mask_vec & (self.0[4] ^ other.0[4]);
    }
}

#[unsafe_target_feature("avx512ifma,avx512vl")]
impl F51x4Reduced {
    #[inline]
    pub fn shuffle(&self, control: Shuffle) -> F51x4Reduced {
        F51x4Reduced([
            shuffle_lanes(self.0[0], control),
            shuffle_lanes(self.0[1], control),
            shuffle_lanes(self.0[2], control),
            shuffle_lanes(self.0[3], control),
            shuffle_lanes(self.0[4], control),
        ])
    }

    #[inline]
    pub fn blend(&self, other: &F51x4Reduced, control: Lanes) -> F51x4Reduced {
        F51x4Reduced([
            blend_lanes(self.0[0], other.0[0], control),
            blend_lanes(self.0[1], other.0[1], control),
            blend_lanes(self.0[2], other.0[2], control),
            blend_lanes(self.0[3], other.0[3], control),
            blend_lanes(self.0[4], other.0[4], control),
        ])
    }

    #[inline]
    pub fn square(&self) -> F51x4Unreduced {
        unsafe {
            let x = &self.0;

            // Represent values with coeff. 2
            let mut z0_2 = u64x4::splat(0);
            let mut z1_2 = u64x4::splat(0);
            let mut z2_2 = u64x4::splat(0);
            let mut z3_2 = u64x4::splat(0);
            let mut z4_2 = u64x4::splat(0);
            let mut z5_2 = u64x4::splat(0);
            let mut z6_2 = u64x4::splat(0);
            let mut z7_2 = u64x4::splat(0);
            let mut z9_2 = u64x4::splat(0);

            // Represent values with coeff. 4
            let mut z2_4 = u64x4::splat(0);
            let mut z3_4 = u64x4::splat(0);
            let mut z4_4 = u64x4::splat(0);
            let mut z5_4 = u64x4::splat(0);
            let mut z6_4 = u64x4::splat(0);
            let mut z7_4 = u64x4::splat(0);
            let mut z8_4 = u64x4::splat(0);

            let mut z0_1 = u64x4::splat(0);
            z0_1 = madd52lo(z0_1, x[0], x[0]);

            let mut z1_1 = u64x4::splat(0);
            z1_2 = madd52lo(z1_2, x[0], x[1]);
            z1_2 = madd52hi(z1_2, x[0], x[0]);

            z2_4 = madd52hi(z2_4, x[0], x[1]);
            let mut z2_1 = z2_4.shl::<2>();
            z2_2 = madd52lo(z2_2, x[0], x[2]);
            z2_1 = madd52lo(z2_1, x[1], x[1]);

            z3_4 = madd52hi(z3_4, x[0], x[2]);
            let mut z3_1 = z3_4.shl::<2>();
            z3_2 = madd52lo(z3_2, x[1], x[2]);
            z3_2 = madd52lo(z3_2, x[0], x[3]);
            z3_2 = madd52hi(z3_2, x[1], x[1]);

            z4_4 = madd52hi(z4_4, x[1], x[2]);
            z4_4 = madd52hi(z4_4, x[0], x[3]);
            let mut z4_1 = z4_4.shl::<2>();
            z4_2 = madd52lo(z4_2, x[1], x[3]);
            z4_2 = madd52lo(z4_2, x[0], x[4]);
            z4_1 = madd52lo(z4_1, x[2], x[2]);

            z5_4 = madd52hi(z5_4, x[1], x[3]);
            z5_4 = madd52hi(z5_4, x[0], x[4]);
            let mut z5_1 = z5_4.shl::<2>();
            z5_2 = madd52lo(z5_2, x[2], x[3]);
            z5_2 = madd52lo(z5_2, x[1], x[4]);
            z5_2 = madd52hi(z5_2, x[2], x[2]);

            z6_4 = madd52hi(z6_4, x[2], x[3]);
            z6_4 = madd52hi(z6_4, x[1], x[4]);
            let mut z6_1 = z6_4.shl::<2>();
            z6_2 = madd52lo(z6_2, x[2], x[4]);
            z6_1 = madd52lo(z6_1, x[3], x[3]);

            z7_4 = madd52hi(z7_4, x[2], x[4]);
            let mut z7_1 = z7_4.shl::<2>();
            z7_2 = madd52lo(z7_2, x[3], x[4]);
            z7_2 = madd52hi(z7_2, x[3], x[3]);

            z8_4 = madd52hi(z8_4, x[3], x[4]);
            let mut z8_1 = z8_4.shl::<2>();
            z8_1 = madd52lo(z8_1, x[4], x[4]);

            let mut z9_1 = u64x4::splat(0);
            z9_2 = madd52hi(z9_2, x[4], x[4]);

            z5_1 += z5_2.shl::<1>();
            z6_1 += z6_2.shl::<1>();
            z7_1 += z7_2.shl::<1>();
            z9_1 += z9_2.shl::<1>();

            let mut t0 = u64x4::splat(0);
            let mut t1 = u64x4::splat(0);
            let r19 = u64x4::splat(19);

            t0 = madd52hi(t0, r19, z9_1);
            t1 = madd52lo(t1, r19, z9_1.shr::<52>());

            z4_2 = madd52lo(z4_2, r19, z8_1.shr::<52>());
            z3_2 = madd52lo(z3_2, r19, z7_1.shr::<52>());
            z2_2 = madd52lo(z2_2, r19, z6_1.shr::<52>());
            z1_2 = madd52lo(z1_2, r19, z5_1.shr::<52>());

            z0_2 = madd52lo(z0_2, r19, t0 + t1);
            z1_2 = madd52hi(z1_2, r19, z5_1);
            z2_2 = madd52hi(z2_2, r19, z6_1);
            z3_2 = madd52hi(z3_2, r19, z7_1);
            z4_2 = madd52hi(z4_2, r19, z8_1);

            z0_1 = madd52lo(z0_1, r19, z5_1);
            z1_1 = madd52lo(z1_1, r19, z6_1);
            z2_1 = madd52lo(z2_1, r19, z7_1);
            z3_1 = madd52lo(z3_1, r19, z8_1);
            z4_1 = madd52lo(z4_1, r19, z9_1);

            F51x4Unreduced([
                z0_1 + z0_2 + z0_2,
                z1_1 + z1_2 + z1_2,
                z2_1 + z2_2 + z2_2,
                z3_1 + z3_2 + z3_2,
                z4_1 + z4_2 + z4_2,
            ])
        }
    }
}

#[unsafe_target_feature("avx512ifma,avx512vl")]
impl From<F51x4Reduced> for F51x4Unreduced {
    #[inline]
    fn from(x: F51x4Reduced) -> F51x4Unreduced {
        F51x4Unreduced(x.0)
    }
}

#[unsafe_target_feature("avx512ifma,avx512vl")]
impl From<F51x4Unreduced> for F51x4Reduced {
    #[inline]
    fn from(x: F51x4Unreduced) -> F51x4Reduced {
        let mask = u64x4::splat((1 << 51) - 1);
        let r19 = u64x4::splat(19);

        // Compute carryouts in parallel
        let c0 = x.0[0].shr::<51>();
        let c1 = x.0[1].shr::<51>();
        let c2 = x.0[2].shr::<51>();
        let c3 = x.0[3].shr::<51>();
        let c4 = x.0[4].shr::<51>();

        unsafe {
            F51x4Reduced([
                madd52lo(x.0[0] & mask, c4, r19),
                (x.0[1] & mask) + c0,
                (x.0[2] & mask) + c1,
                (x.0[3] & mask) + c2,
                (x.0[4] & mask) + c3,
            ])
        }
    }
}

#[unsafe_target_feature("avx512ifma,avx512vl")]
impl Add<F51x4Unreduced> for F51x4Unreduced {
    type Output = F51x4Unreduced;
    #[inline]
    fn add(self, rhs: F51x4Unreduced) -> F51x4Unreduced {
        F51x4Unreduced([
            self.0[0] + rhs.0[0],
            self.0[1] + rhs.0[1],
            self.0[2] + rhs.0[2],
            self.0[3] + rhs.0[3],
            self.0[4] + rhs.0[4],
        ])
    }
}

#[unsafe_target_feature("avx512ifma,avx512vl")]
impl<'a> Mul<(u32, u32, u32, u32)> for &'a F51x4Reduced {
    type Output = F51x4Unreduced;
    #[inline]
    fn mul(self, scalars: (u32, u32, u32, u32)) -> F51x4Unreduced {
        unsafe {
            let x = &self.0;
            let y = u64x4::new(
                scalars.0 as u64,
                scalars.1 as u64,
                scalars.2 as u64,
                scalars.3 as u64,
            );
            let r19 = u64x4::splat(19);

            let mut z0_1 = u64x4::splat(0);
            let mut z1_1 = u64x4::splat(0);
            let mut z2_1 = u64x4::splat(0);
            let mut z3_1 = u64x4::splat(0);
            let mut z4_1 = u64x4::splat(0);
            let mut z1_2 = u64x4::splat(0);
            let mut z2_2 = u64x4::splat(0);
            let mut z3_2 = u64x4::splat(0);
            let mut z4_2 = u64x4::splat(0);
            let mut z5_2 = u64x4::splat(0);

            // Wave 0
            z4_2 = madd52hi(z4_2, y, x[3]);
            z5_2 = madd52hi(z5_2, y, x[4]);
            z4_1 = madd52lo(z4_1, y, x[4]);
            z0_1 = madd52lo(z0_1, y, x[0]);
            z3_1 = madd52lo(z3_1, y, x[3]);
            z2_1 = madd52lo(z2_1, y, x[2]);
            z1_1 = madd52lo(z1_1, y, x[1]);
            z3_2 = madd52hi(z3_2, y, x[2]);

            // Wave 2
            z2_2 = madd52hi(z2_2, y, x[1]);
            z1_2 = madd52hi(z1_2, y, x[0]);
            z0_1 = madd52lo(z0_1, z5_2 + z5_2, r19);

            F51x4Unreduced([
                z0_1,
                z1_1 + z1_2 + z1_2,
                z2_1 + z2_2 + z2_2,
                z3_1 + z3_2 + z3_2,
                z4_1 + z4_2 + z4_2,
            ])
        }
    }
}

#[unsafe_target_feature("avx512ifma,avx512vl")]
impl<'a, 'b> Mul<&'b F51x4Reduced> for &'a F51x4Reduced {
    type Output = F51x4Unreduced;
    #[inline]
    fn mul(self, rhs: &'b F51x4Reduced) -> F51x4Unreduced {
        unsafe {
            // Inputs
            let x = &self.0;
            let y = &rhs.0;

            // Accumulators for terms with coeff 1
            let mut z0_1 = u64x4::splat(0);
            let mut z1_1 = u64x4::splat(0);
            let mut z2_1 = u64x4::splat(0);
            let mut z3_1 = u64x4::splat(0);
            let mut z4_1 = u64x4::splat(0);
            let mut z5_1 = u64x4::splat(0);
            let mut z6_1 = u64x4::splat(0);
            let mut z7_1 = u64x4::splat(0);
            let mut z8_1 = u64x4::splat(0);

            // Accumulators for terms with coeff 2
            let mut z0_2 = u64x4::splat(0);
            let mut z1_2 = u64x4::splat(0);
            let mut z2_2 = u64x4::splat(0);
            let mut z3_2 = u64x4::splat(0);
            let mut z4_2 = u64x4::splat(0);
            let mut z5_2 = u64x4::splat(0);
            let mut z6_2 = u64x4::splat(0);
            let mut z7_2 = u64x4::splat(0);
            let mut z8_2 = u64x4::splat(0);
            let mut z9_2 = u64x4::splat(0);

            // LLVM doesn't seem to do much work reordering IFMA
            // instructions, so try to organize them into "waves" of 8
            // independent operations (4c latency, 0.5 c throughput
            // means 8 in flight)

            // Wave 0
            z4_1 = madd52lo(z4_1, x[2], y[2]);
            z5_2 = madd52hi(z5_2, x[2], y[2]);
            z5_1 = madd52lo(z5_1, x[4], y[1]);
            z6_2 = madd52hi(z6_2, x[4], y[1]);
            z6_1 = madd52lo(z6_1, x[4], y[2]);
            z7_2 = madd52hi(z7_2, x[4], y[2]);
            z7_1 = madd52lo(z7_1, x[4], y[3]);
            z8_2 = madd52hi(z8_2, x[4], y[3]);

            // Wave 1
            z4_1 = madd52lo(z4_1, x[3], y[1]);
            z5_2 = madd52hi(z5_2, x[3], y[1]);
            z5_1 = madd52lo(z5_1, x[3], y[2]);
            z6_2 = madd52hi(z6_2, x[3], y[2]);
            z6_1 = madd52lo(z6_1, x[3], y[3]);
            z7_2 = madd52hi(z7_2, x[3], y[3]);
            z7_1 = madd52lo(z7_1, x[3], y[4]);
            z8_2 = madd52hi(z8_2, x[3], y[4]);

            // Wave 2
            z8_1 = madd52lo(z8_1, x[4], y[4]);
            z9_2 = madd52hi(z9_2, x[4], y[4]);
            z4_1 = madd52lo(z4_1, x[4], y[0]);
            z5_2 = madd52hi(z5_2, x[4], y[0]);
            z5_1 = madd52lo(z5_1, x[2], y[3]);
            z6_2 = madd52hi(z6_2, x[2], y[3]);
            z6_1 = madd52lo(z6_1, x[2], y[4]);
            z7_2 = madd52hi(z7_2, x[2], y[4]);

            let z8 = z8_1 + z8_2 + z8_2;
            let z9 = z9_2 + z9_2;

            // Wave 3
            z3_1 = madd52lo(z3_1, x[3], y[0]);
            z4_2 = madd52hi(z4_2, x[3], y[0]);
            z4_1 = madd52lo(z4_1, x[1], y[3]);
            z5_2 = madd52hi(z5_2, x[1], y[3]);
            z5_1 = madd52lo(z5_1, x[1], y[4]);
            z6_2 = madd52hi(z6_2, x[1], y[4]);
            z2_1 = madd52lo(z2_1, x[2], y[0]);
            z3_2 = madd52hi(z3_2, x[2], y[0]);

            let z6 = z6_1 + z6_2 + z6_2;
            let z7 = z7_1 + z7_2 + z7_2;

            // Wave 4
            z3_1 = madd52lo(z3_1, x[2], y[1]);
            z4_2 = madd52hi(z4_2, x[2], y[1]);
            z4_1 = madd52lo(z4_1, x[0], y[4]);
            z5_2 = madd52hi(z5_2, x[0], y[4]);
            z1_1 = madd52lo(z1_1, x[1], y[0]);
            z2_2 = madd52hi(z2_2, x[1], y[0]);
            z2_1 = madd52lo(z2_1, x[1], y[1]);
            z3_2 = madd52hi(z3_2, x[1], y[1]);

            let z5 = z5_1 + z5_2 + z5_2;

            // Wave 5
            z3_1 = madd52lo(z3_1, x[1], y[2]);
            z4_2 = madd52hi(z4_2, x[1], y[2]);
            z0_1 = madd52lo(z0_1, x[0], y[0]);
            z1_2 = madd52hi(z1_2, x[0], y[0]);
            z1_1 = madd52lo(z1_1, x[0], y[1]);
            z2_1 = madd52lo(z2_1, x[0], y[2]);
            z2_2 = madd52hi(z2_2, x[0], y[1]);
            z3_2 = madd52hi(z3_2, x[0], y[2]);

            let mut t0 = u64x4::splat(0);
            let mut t1 = u64x4::splat(0);
            let r19 = u64x4::splat(19);

            // Wave 6
            t0 = madd52hi(t0, r19, z9);
            t1 = madd52lo(t1, r19, z9.shr::<52>());
            z3_1 = madd52lo(z3_1, x[0], y[3]);
            z4_2 = madd52hi(z4_2, x[0], y[3]);
            z1_2 = madd52lo(z1_2, r19, z5.shr::<52>());
            z2_2 = madd52lo(z2_2, r19, z6.shr::<52>());
            z3_2 = madd52lo(z3_2, r19, z7.shr::<52>());
            z0_1 = madd52lo(z0_1, r19, z5);

            // Wave 7
            z4_1 = madd52lo(z4_1, r19, z9);
            z1_1 = madd52lo(z1_1, r19, z6);
            z0_2 = madd52lo(z0_2, r19, t0 + t1);
            z4_2 = madd52hi(z4_2, r19, z8);
            z2_1 = madd52lo(z2_1, r19, z7);
            z1_2 = madd52hi(z1_2, r19, z5);
            z2_2 = madd52hi(z2_2, r19, z6);
            z3_2 = madd52hi(z3_2, r19, z7);

            // Wave 8
            z3_1 = madd52lo(z3_1, r19, z8);
            z4_2 = madd52lo(z4_2, r19, z8.shr::<52>());

            F51x4Unreduced([
                z0_1 + z0_2 + z0_2,
                z1_1 + z1_2 + z1_2,
                z2_1 + z2_2 + z2_2,
                z3_1 + z3_2 + z3_2,
                z4_1 + z4_2 + z4_2,
            ])
        }
    }
}

#[cfg(target_feature = "avx512ifma,avx512vl")]
#[cfg(test)]
mod test {
    use super::*;

    #[test]
    fn vpmadd52luq() {
        let x = u64x4::splat(2);
        let y = u64x4::splat(3);
        let mut z = u64x4::splat(5);

        z = unsafe { madd52lo(z, x, y) };

        assert_eq!(z, u64x4::splat(5 + 2 * 3));
    }

    #[test]
    fn new_split_round_trip_on_reduced_input() {
        // Invert a small field element to get a big one
        let a = FieldElement51([2438, 24, 243, 0, 0]).invert();

        let ax4 = F51x4Unreduced::new(&a, &a, &a, &a);
        let splits = ax4.split();

        for i in 0..4 {
            assert_eq!(a, splits[i]);
        }
    }

    #[test]
    fn new_split_round_trip_on_unreduced_input() {
        // Invert a small field element to get a big one
        let a = FieldElement51([2438, 24, 243, 0, 0]).invert();
        // ... but now multiply it by 16 without reducing coeffs
        let a16 = FieldElement51([
            a.0[0] << 4,
            a.0[1] << 4,
            a.0[2] << 4,
            a.0[3] << 4,
            a.0[4] << 4,
        ]);

        let a16x4 = F51x4Unreduced::new(&a16, &a16, &a16, &a16);
        let splits = a16x4.split();

        for i in 0..4 {
            assert_eq!(a16, splits[i]);
        }
    }

    #[test]
    fn test_reduction() {
        // Invert a small field element to get a big one
        let a = FieldElement51([2438, 24, 243, 0, 0]).invert();
        // ... but now multiply it by 128 without reducing coeffs
        let abig = FieldElement51([
            a.0[0] << 4,
            a.0[1] << 4,
            a.0[2] << 4,
            a.0[3] << 4,
            a.0[4] << 4,
        ]);

        let abigx4: F51x4Reduced = F51x4Unreduced::new(&abig, &abig, &abig, &abig).into();

        let splits = F51x4Unreduced::from(abigx4).split();
        let c = &a * &FieldElement51([(1 << 4), 0, 0, 0, 0]);

        for i in 0..4 {
            assert_eq!(c, splits[i]);
        }
    }

    #[test]
    fn mul_matches_serial() {
        // Invert a small field element to get a big one
        let a = FieldElement51([2438, 24, 243, 0, 0]).invert();
        let b = FieldElement51([98098, 87987897, 0, 1, 0]).invert();
        let c = &a * &b;

        let ax4: F51x4Reduced = F51x4Unreduced::new(&a, &a, &a, &a).into();
        let bx4: F51x4Reduced = F51x4Unreduced::new(&b, &b, &b, &b).into();
        let cx4 = &ax4 * &bx4;

        let splits = cx4.split();

        for i in 0..4 {
            assert_eq!(c, splits[i]);
        }
    }

    #[test]
    fn iterated_mul_matches_serial() {
        // Invert a small field element to get a big one
        let a = FieldElement51([2438, 24, 243, 0, 0]).invert();
        let b = FieldElement51([98098, 87987897, 0, 1, 0]).invert();
        let mut c = &a * &b;
        for _i in 0..1024 {
            c = &a * &c;
            c = &b * &c;
        }

        let ax4: F51x4Reduced = F51x4Unreduced::new(&a, &a, &a, &a).into();
        let bx4: F51x4Reduced = F51x4Unreduced::new(&b, &b, &b, &b).into();
        let mut cx4 = &ax4 * &bx4;
        for _i in 0..1024 {
            cx4 = &ax4 * &F51x4Reduced::from(cx4);
            cx4 = &bx4 * &F51x4Reduced::from(cx4);
        }

        let splits = cx4.split();

        for i in 0..4 {
            assert_eq!(c, splits[i]);
        }
    }

    #[test]
    fn square_matches_mul() {
        // Invert a small field element to get a big one
        let a = FieldElement51([2438, 24, 243, 0, 0]).invert();

        let ax4: F51x4Reduced = F51x4Unreduced::new(&a, &a, &a, &a).into();
        let cx4 = &ax4 * &ax4;
        let cx4_sq = ax4.square();

        let splits = cx4.split();
        let splits_sq = cx4_sq.split();

        for i in 0..4 {
            assert_eq!(splits_sq[i], splits[i]);
        }
    }

    #[test]
    fn iterated_square_matches_serial() {
        // Invert a small field element to get a big one
        let mut a = FieldElement51([2438, 24, 243, 0, 0]).invert();
        let mut ax4 = F51x4Unreduced::new(&a, &a, &a, &a);
        for _j in 0..1024 {
            a = a.square();
            ax4 = F51x4Reduced::from(ax4).square();

            let splits = ax4.split();
            for i in 0..4 {
                assert_eq!(a, splits[i]);
            }
        }
    }

    #[test]
    fn iterated_u32_mul_matches_serial() {
        // Invert a small field element to get a big one
        let a = FieldElement51([2438, 24, 243, 0, 0]).invert();
        let b = FieldElement51([121665, 0, 0, 0, 0]);
        let mut c = &a * &b;
        for _i in 0..1024 {
            c = &b * &c;
        }

        let ax4 = F51x4Unreduced::new(&a, &a, &a, &a);
        let bx4 = (121665u32, 121665u32, 121665u32, 121665u32);
        let mut cx4 = &F51x4Reduced::from(ax4) * bx4;
        for _i in 0..1024 {
            cx4 = &F51x4Reduced::from(cx4) * bx4;
        }

        let splits = cx4.split();

        for i in 0..4 {
            assert_eq!(c, splits[i]);
        }
    }

    #[test]
    fn shuffle_AAAA() {
        let x0 = FieldElement51::from_bytes(&[0x10; 32]);
        let x1 = FieldElement51::from_bytes(&[0x11; 32]);
        let x2 = FieldElement51::from_bytes(&[0x12; 32]);
        let x3 = FieldElement51::from_bytes(&[0x13; 32]);

        let x = F51x4Unreduced::new(&x0, &x1, &x2, &x3);

        let y = x.shuffle(Shuffle::AAAA);
        let splits = y.split();

        assert_eq!(splits[0], x0);
        assert_eq!(splits[1], x0);
        assert_eq!(splits[2], x0);
        assert_eq!(splits[3], x0);
    }

    #[test]
    fn blend_AB() {
        let x0 = FieldElement51::from_bytes(&[0x10; 32]);
        let x1 = FieldElement51::from_bytes(&[0x11; 32]);
        let x2 = FieldElement51::from_bytes(&[0x12; 32]);
        let x3 = FieldElement51::from_bytes(&[0x13; 32]);

        let x = F51x4Unreduced::new(&x0, &x1, &x2, &x3);
        let z = F51x4Unreduced::new(&x3, &x2, &x1, &x0);

        let y = x.blend(&z, Lanes::AB);
        let splits = y.split();

        assert_eq!(splits[0], x3);
        assert_eq!(splits[1], x2);
        assert_eq!(splits[2], x2);
        assert_eq!(splits[3], x3);
    }
}