@@ -531,160 +531,80 @@ define void @test(ptr %A, ptr noalias %B) #0 {
531
531
; MAX-BW-NEXT: [[TMP13:%.*]] = add i64 [[OFFSET_IDX]], 26
532
532
; MAX-BW-NEXT: [[TMP14:%.*]] = add i64 [[OFFSET_IDX]], 28
533
533
; MAX-BW-NEXT: [[TMP15:%.*]] = add i64 [[OFFSET_IDX]], 30
534
- ; MAX-BW-NEXT: [[TMP16:%.*]] = add i64 [[OFFSET_IDX]], 32
535
- ; MAX-BW-NEXT: [[TMP17:%.*]] = add i64 [[OFFSET_IDX]], 34
536
- ; MAX-BW-NEXT: [[TMP18:%.*]] = add i64 [[OFFSET_IDX]], 36
537
- ; MAX-BW-NEXT: [[TMP19:%.*]] = add i64 [[OFFSET_IDX]], 38
538
- ; MAX-BW-NEXT: [[TMP20:%.*]] = add i64 [[OFFSET_IDX]], 40
539
- ; MAX-BW-NEXT: [[TMP21:%.*]] = add i64 [[OFFSET_IDX]], 42
540
- ; MAX-BW-NEXT: [[TMP22:%.*]] = add i64 [[OFFSET_IDX]], 44
541
- ; MAX-BW-NEXT: [[TMP23:%.*]] = add i64 [[OFFSET_IDX]], 46
542
- ; MAX-BW-NEXT: [[TMP24:%.*]] = add i64 [[OFFSET_IDX]], 48
543
- ; MAX-BW-NEXT: [[TMP25:%.*]] = add i64 [[OFFSET_IDX]], 50
544
- ; MAX-BW-NEXT: [[TMP26:%.*]] = add i64 [[OFFSET_IDX]], 52
545
- ; MAX-BW-NEXT: [[TMP27:%.*]] = add i64 [[OFFSET_IDX]], 54
546
- ; MAX-BW-NEXT: [[TMP28:%.*]] = add i64 [[OFFSET_IDX]], 56
547
- ; MAX-BW-NEXT: [[TMP29:%.*]] = add i64 [[OFFSET_IDX]], 58
548
- ; MAX-BW-NEXT: [[TMP30:%.*]] = add i64 [[OFFSET_IDX]], 60
549
- ; MAX-BW-NEXT: [[TMP31:%.*]] = add i64 [[OFFSET_IDX]], 62
550
- ; MAX-BW-NEXT: [[TMP32:%.*]] = add nuw nsw i64 [[TMP0]], 0
551
- ; MAX-BW-NEXT: [[TMP33:%.*]] = add nuw nsw i64 [[TMP1]], 0
552
- ; MAX-BW-NEXT: [[TMP34:%.*]] = add nuw nsw i64 [[TMP2]], 0
553
- ; MAX-BW-NEXT: [[TMP35:%.*]] = add nuw nsw i64 [[TMP3]], 0
554
- ; MAX-BW-NEXT: [[TMP36:%.*]] = add nuw nsw i64 [[TMP4]], 0
555
- ; MAX-BW-NEXT: [[TMP37:%.*]] = add nuw nsw i64 [[TMP5]], 0
556
- ; MAX-BW-NEXT: [[TMP38:%.*]] = add nuw nsw i64 [[TMP6]], 0
557
- ; MAX-BW-NEXT: [[TMP39:%.*]] = add nuw nsw i64 [[TMP7]], 0
558
- ; MAX-BW-NEXT: [[TMP40:%.*]] = add nuw nsw i64 [[TMP8]], 0
559
- ; MAX-BW-NEXT: [[TMP41:%.*]] = add nuw nsw i64 [[TMP9]], 0
560
- ; MAX-BW-NEXT: [[TMP42:%.*]] = add nuw nsw i64 [[TMP10]], 0
561
- ; MAX-BW-NEXT: [[TMP43:%.*]] = add nuw nsw i64 [[TMP11]], 0
562
- ; MAX-BW-NEXT: [[TMP44:%.*]] = add nuw nsw i64 [[TMP12]], 0
563
- ; MAX-BW-NEXT: [[TMP45:%.*]] = add nuw nsw i64 [[TMP13]], 0
564
- ; MAX-BW-NEXT: [[TMP46:%.*]] = add nuw nsw i64 [[TMP14]], 0
565
- ; MAX-BW-NEXT: [[TMP47:%.*]] = add nuw nsw i64 [[TMP15]], 0
566
- ; MAX-BW-NEXT: [[TMP48:%.*]] = add nuw nsw i64 [[TMP16]], 0
567
- ; MAX-BW-NEXT: [[TMP49:%.*]] = add nuw nsw i64 [[TMP17]], 0
568
- ; MAX-BW-NEXT: [[TMP50:%.*]] = add nuw nsw i64 [[TMP18]], 0
569
- ; MAX-BW-NEXT: [[TMP51:%.*]] = add nuw nsw i64 [[TMP19]], 0
570
- ; MAX-BW-NEXT: [[TMP52:%.*]] = add nuw nsw i64 [[TMP20]], 0
571
- ; MAX-BW-NEXT: [[TMP53:%.*]] = add nuw nsw i64 [[TMP21]], 0
572
- ; MAX-BW-NEXT: [[TMP54:%.*]] = add nuw nsw i64 [[TMP22]], 0
573
- ; MAX-BW-NEXT: [[TMP55:%.*]] = add nuw nsw i64 [[TMP23]], 0
574
- ; MAX-BW-NEXT: [[TMP56:%.*]] = add nuw nsw i64 [[TMP24]], 0
575
- ; MAX-BW-NEXT: [[TMP57:%.*]] = add nuw nsw i64 [[TMP25]], 0
576
- ; MAX-BW-NEXT: [[TMP58:%.*]] = add nuw nsw i64 [[TMP26]], 0
577
- ; MAX-BW-NEXT: [[TMP59:%.*]] = add nuw nsw i64 [[TMP27]], 0
578
- ; MAX-BW-NEXT: [[TMP60:%.*]] = add nuw nsw i64 [[TMP28]], 0
579
- ; MAX-BW-NEXT: [[TMP61:%.*]] = add nuw nsw i64 [[TMP29]], 0
580
- ; MAX-BW-NEXT: [[TMP62:%.*]] = add nuw nsw i64 [[TMP30]], 0
581
- ; MAX-BW-NEXT: [[TMP63:%.*]] = add nuw nsw i64 [[TMP31]], 0
582
- ; MAX-BW-NEXT: [[TMP64:%.*]] = getelementptr inbounds [1024 x i32], ptr [[A:%.*]], i64 0, i64 [[TMP32]]
583
- ; MAX-BW-NEXT: [[TMP65:%.*]] = getelementptr inbounds i32, ptr [[TMP64]], i32 0
584
- ; MAX-BW-NEXT: [[WIDE_VEC:%.*]] = load <64 x i32>, ptr [[TMP65]], align 4
585
- ; MAX-BW-NEXT: [[STRIDED_VEC:%.*]] = shufflevector <64 x i32> [[WIDE_VEC]], <64 x i32> poison, <32 x i32> <i32 0, i32 2, i32 4, i32 6, i32 8, i32 10, i32 12, i32 14, i32 16, i32 18, i32 20, i32 22, i32 24, i32 26, i32 28, i32 30, i32 32, i32 34, i32 36, i32 38, i32 40, i32 42, i32 44, i32 46, i32 48, i32 50, i32 52, i32 54, i32 56, i32 58, i32 60, i32 62>
586
- ; MAX-BW-NEXT: [[STRIDED_VEC1:%.*]] = shufflevector <64 x i32> [[WIDE_VEC]], <64 x i32> poison, <32 x i32> <i32 1, i32 3, i32 5, i32 7, i32 9, i32 11, i32 13, i32 15, i32 17, i32 19, i32 21, i32 23, i32 25, i32 27, i32 29, i32 31, i32 33, i32 35, i32 37, i32 39, i32 41, i32 43, i32 45, i32 47, i32 49, i32 51, i32 53, i32 55, i32 57, i32 59, i32 61, i32 63>
587
- ; MAX-BW-NEXT: [[TMP66:%.*]] = add <32 x i32> [[STRIDED_VEC]], [[STRIDED_VEC1]]
588
- ; MAX-BW-NEXT: [[TMP67:%.*]] = trunc <32 x i32> [[TMP66]] to <32 x i8>
589
- ; MAX-BW-NEXT: [[TMP68:%.*]] = getelementptr inbounds [1024 x i8], ptr [[B:%.*]], i64 0, i64 [[TMP32]]
590
- ; MAX-BW-NEXT: [[TMP69:%.*]] = getelementptr inbounds [1024 x i8], ptr [[B]], i64 0, i64 [[TMP33]]
591
- ; MAX-BW-NEXT: [[TMP70:%.*]] = getelementptr inbounds [1024 x i8], ptr [[B]], i64 0, i64 [[TMP34]]
592
- ; MAX-BW-NEXT: [[TMP71:%.*]] = getelementptr inbounds [1024 x i8], ptr [[B]], i64 0, i64 [[TMP35]]
593
- ; MAX-BW-NEXT: [[TMP72:%.*]] = getelementptr inbounds [1024 x i8], ptr [[B]], i64 0, i64 [[TMP36]]
594
- ; MAX-BW-NEXT: [[TMP73:%.*]] = getelementptr inbounds [1024 x i8], ptr [[B]], i64 0, i64 [[TMP37]]
595
- ; MAX-BW-NEXT: [[TMP74:%.*]] = getelementptr inbounds [1024 x i8], ptr [[B]], i64 0, i64 [[TMP38]]
596
- ; MAX-BW-NEXT: [[TMP75:%.*]] = getelementptr inbounds [1024 x i8], ptr [[B]], i64 0, i64 [[TMP39]]
597
- ; MAX-BW-NEXT: [[TMP76:%.*]] = getelementptr inbounds [1024 x i8], ptr [[B]], i64 0, i64 [[TMP40]]
598
- ; MAX-BW-NEXT: [[TMP77:%.*]] = getelementptr inbounds [1024 x i8], ptr [[B]], i64 0, i64 [[TMP41]]
599
- ; MAX-BW-NEXT: [[TMP78:%.*]] = getelementptr inbounds [1024 x i8], ptr [[B]], i64 0, i64 [[TMP42]]
600
- ; MAX-BW-NEXT: [[TMP79:%.*]] = getelementptr inbounds [1024 x i8], ptr [[B]], i64 0, i64 [[TMP43]]
601
- ; MAX-BW-NEXT: [[TMP80:%.*]] = getelementptr inbounds [1024 x i8], ptr [[B]], i64 0, i64 [[TMP44]]
602
- ; MAX-BW-NEXT: [[TMP81:%.*]] = getelementptr inbounds [1024 x i8], ptr [[B]], i64 0, i64 [[TMP45]]
603
- ; MAX-BW-NEXT: [[TMP82:%.*]] = getelementptr inbounds [1024 x i8], ptr [[B]], i64 0, i64 [[TMP46]]
604
- ; MAX-BW-NEXT: [[TMP83:%.*]] = getelementptr inbounds [1024 x i8], ptr [[B]], i64 0, i64 [[TMP47]]
605
- ; MAX-BW-NEXT: [[TMP84:%.*]] = getelementptr inbounds [1024 x i8], ptr [[B]], i64 0, i64 [[TMP48]]
606
- ; MAX-BW-NEXT: [[TMP85:%.*]] = getelementptr inbounds [1024 x i8], ptr [[B]], i64 0, i64 [[TMP49]]
607
- ; MAX-BW-NEXT: [[TMP86:%.*]] = getelementptr inbounds [1024 x i8], ptr [[B]], i64 0, i64 [[TMP50]]
608
- ; MAX-BW-NEXT: [[TMP87:%.*]] = getelementptr inbounds [1024 x i8], ptr [[B]], i64 0, i64 [[TMP51]]
609
- ; MAX-BW-NEXT: [[TMP88:%.*]] = getelementptr inbounds [1024 x i8], ptr [[B]], i64 0, i64 [[TMP52]]
610
- ; MAX-BW-NEXT: [[TMP89:%.*]] = getelementptr inbounds [1024 x i8], ptr [[B]], i64 0, i64 [[TMP53]]
611
- ; MAX-BW-NEXT: [[TMP90:%.*]] = getelementptr inbounds [1024 x i8], ptr [[B]], i64 0, i64 [[TMP54]]
612
- ; MAX-BW-NEXT: [[TMP91:%.*]] = getelementptr inbounds [1024 x i8], ptr [[B]], i64 0, i64 [[TMP55]]
613
- ; MAX-BW-NEXT: [[TMP92:%.*]] = getelementptr inbounds [1024 x i8], ptr [[B]], i64 0, i64 [[TMP56]]
614
- ; MAX-BW-NEXT: [[TMP93:%.*]] = getelementptr inbounds [1024 x i8], ptr [[B]], i64 0, i64 [[TMP57]]
615
- ; MAX-BW-NEXT: [[TMP94:%.*]] = getelementptr inbounds [1024 x i8], ptr [[B]], i64 0, i64 [[TMP58]]
616
- ; MAX-BW-NEXT: [[TMP95:%.*]] = getelementptr inbounds [1024 x i8], ptr [[B]], i64 0, i64 [[TMP59]]
617
- ; MAX-BW-NEXT: [[TMP96:%.*]] = getelementptr inbounds [1024 x i8], ptr [[B]], i64 0, i64 [[TMP60]]
618
- ; MAX-BW-NEXT: [[TMP97:%.*]] = getelementptr inbounds [1024 x i8], ptr [[B]], i64 0, i64 [[TMP61]]
619
- ; MAX-BW-NEXT: [[TMP98:%.*]] = getelementptr inbounds [1024 x i8], ptr [[B]], i64 0, i64 [[TMP62]]
620
- ; MAX-BW-NEXT: [[TMP99:%.*]] = getelementptr inbounds [1024 x i8], ptr [[B]], i64 0, i64 [[TMP63]]
621
- ; MAX-BW-NEXT: [[TMP100:%.*]] = extractelement <32 x i8> [[TMP67]], i32 0
622
- ; MAX-BW-NEXT: store i8 [[TMP100]], ptr [[TMP68]], align 1
623
- ; MAX-BW-NEXT: [[TMP101:%.*]] = extractelement <32 x i8> [[TMP67]], i32 1
624
- ; MAX-BW-NEXT: store i8 [[TMP101]], ptr [[TMP69]], align 1
625
- ; MAX-BW-NEXT: [[TMP102:%.*]] = extractelement <32 x i8> [[TMP67]], i32 2
626
- ; MAX-BW-NEXT: store i8 [[TMP102]], ptr [[TMP70]], align 1
627
- ; MAX-BW-NEXT: [[TMP103:%.*]] = extractelement <32 x i8> [[TMP67]], i32 3
628
- ; MAX-BW-NEXT: store i8 [[TMP103]], ptr [[TMP71]], align 1
629
- ; MAX-BW-NEXT: [[TMP104:%.*]] = extractelement <32 x i8> [[TMP67]], i32 4
630
- ; MAX-BW-NEXT: store i8 [[TMP104]], ptr [[TMP72]], align 1
631
- ; MAX-BW-NEXT: [[TMP105:%.*]] = extractelement <32 x i8> [[TMP67]], i32 5
632
- ; MAX-BW-NEXT: store i8 [[TMP105]], ptr [[TMP73]], align 1
633
- ; MAX-BW-NEXT: [[TMP106:%.*]] = extractelement <32 x i8> [[TMP67]], i32 6
634
- ; MAX-BW-NEXT: store i8 [[TMP106]], ptr [[TMP74]], align 1
635
- ; MAX-BW-NEXT: [[TMP107:%.*]] = extractelement <32 x i8> [[TMP67]], i32 7
636
- ; MAX-BW-NEXT: store i8 [[TMP107]], ptr [[TMP75]], align 1
637
- ; MAX-BW-NEXT: [[TMP108:%.*]] = extractelement <32 x i8> [[TMP67]], i32 8
638
- ; MAX-BW-NEXT: store i8 [[TMP108]], ptr [[TMP76]], align 1
639
- ; MAX-BW-NEXT: [[TMP109:%.*]] = extractelement <32 x i8> [[TMP67]], i32 9
640
- ; MAX-BW-NEXT: store i8 [[TMP109]], ptr [[TMP77]], align 1
641
- ; MAX-BW-NEXT: [[TMP110:%.*]] = extractelement <32 x i8> [[TMP67]], i32 10
642
- ; MAX-BW-NEXT: store i8 [[TMP110]], ptr [[TMP78]], align 1
643
- ; MAX-BW-NEXT: [[TMP111:%.*]] = extractelement <32 x i8> [[TMP67]], i32 11
644
- ; MAX-BW-NEXT: store i8 [[TMP111]], ptr [[TMP79]], align 1
645
- ; MAX-BW-NEXT: [[TMP112:%.*]] = extractelement <32 x i8> [[TMP67]], i32 12
646
- ; MAX-BW-NEXT: store i8 [[TMP112]], ptr [[TMP80]], align 1
647
- ; MAX-BW-NEXT: [[TMP113:%.*]] = extractelement <32 x i8> [[TMP67]], i32 13
648
- ; MAX-BW-NEXT: store i8 [[TMP113]], ptr [[TMP81]], align 1
649
- ; MAX-BW-NEXT: [[TMP114:%.*]] = extractelement <32 x i8> [[TMP67]], i32 14
650
- ; MAX-BW-NEXT: store i8 [[TMP114]], ptr [[TMP82]], align 1
651
- ; MAX-BW-NEXT: [[TMP115:%.*]] = extractelement <32 x i8> [[TMP67]], i32 15
652
- ; MAX-BW-NEXT: store i8 [[TMP115]], ptr [[TMP83]], align 1
653
- ; MAX-BW-NEXT: [[TMP116:%.*]] = extractelement <32 x i8> [[TMP67]], i32 16
654
- ; MAX-BW-NEXT: store i8 [[TMP116]], ptr [[TMP84]], align 1
655
- ; MAX-BW-NEXT: [[TMP117:%.*]] = extractelement <32 x i8> [[TMP67]], i32 17
656
- ; MAX-BW-NEXT: store i8 [[TMP117]], ptr [[TMP85]], align 1
657
- ; MAX-BW-NEXT: [[TMP118:%.*]] = extractelement <32 x i8> [[TMP67]], i32 18
658
- ; MAX-BW-NEXT: store i8 [[TMP118]], ptr [[TMP86]], align 1
659
- ; MAX-BW-NEXT: [[TMP119:%.*]] = extractelement <32 x i8> [[TMP67]], i32 19
660
- ; MAX-BW-NEXT: store i8 [[TMP119]], ptr [[TMP87]], align 1
661
- ; MAX-BW-NEXT: [[TMP120:%.*]] = extractelement <32 x i8> [[TMP67]], i32 20
662
- ; MAX-BW-NEXT: store i8 [[TMP120]], ptr [[TMP88]], align 1
663
- ; MAX-BW-NEXT: [[TMP121:%.*]] = extractelement <32 x i8> [[TMP67]], i32 21
664
- ; MAX-BW-NEXT: store i8 [[TMP121]], ptr [[TMP89]], align 1
665
- ; MAX-BW-NEXT: [[TMP122:%.*]] = extractelement <32 x i8> [[TMP67]], i32 22
666
- ; MAX-BW-NEXT: store i8 [[TMP122]], ptr [[TMP90]], align 1
667
- ; MAX-BW-NEXT: [[TMP123:%.*]] = extractelement <32 x i8> [[TMP67]], i32 23
668
- ; MAX-BW-NEXT: store i8 [[TMP123]], ptr [[TMP91]], align 1
669
- ; MAX-BW-NEXT: [[TMP124:%.*]] = extractelement <32 x i8> [[TMP67]], i32 24
670
- ; MAX-BW-NEXT: store i8 [[TMP124]], ptr [[TMP92]], align 1
671
- ; MAX-BW-NEXT: [[TMP125:%.*]] = extractelement <32 x i8> [[TMP67]], i32 25
672
- ; MAX-BW-NEXT: store i8 [[TMP125]], ptr [[TMP93]], align 1
673
- ; MAX-BW-NEXT: [[TMP126:%.*]] = extractelement <32 x i8> [[TMP67]], i32 26
674
- ; MAX-BW-NEXT: store i8 [[TMP126]], ptr [[TMP94]], align 1
675
- ; MAX-BW-NEXT: [[TMP127:%.*]] = extractelement <32 x i8> [[TMP67]], i32 27
676
- ; MAX-BW-NEXT: store i8 [[TMP127]], ptr [[TMP95]], align 1
677
- ; MAX-BW-NEXT: [[TMP128:%.*]] = extractelement <32 x i8> [[TMP67]], i32 28
678
- ; MAX-BW-NEXT: store i8 [[TMP128]], ptr [[TMP96]], align 1
679
- ; MAX-BW-NEXT: [[TMP129:%.*]] = extractelement <32 x i8> [[TMP67]], i32 29
680
- ; MAX-BW-NEXT: store i8 [[TMP129]], ptr [[TMP97]], align 1
681
- ; MAX-BW-NEXT: [[TMP130:%.*]] = extractelement <32 x i8> [[TMP67]], i32 30
682
- ; MAX-BW-NEXT: store i8 [[TMP130]], ptr [[TMP98]], align 1
683
- ; MAX-BW-NEXT: [[TMP131:%.*]] = extractelement <32 x i8> [[TMP67]], i32 31
684
- ; MAX-BW-NEXT: store i8 [[TMP131]], ptr [[TMP99]], align 1
685
- ; MAX-BW-NEXT: [[INDEX_NEXT]] = add nuw i64 [[INDEX]], 32
686
- ; MAX-BW-NEXT: [[TMP132:%.*]] = icmp eq i64 [[INDEX_NEXT]], 512
687
- ; MAX-BW-NEXT: br i1 [[TMP132]], label [[MIDDLE_BLOCK:%.*]], label [[VECTOR_BODY]], !llvm.loop [[LOOP9:![0-9]+]]
534
+ ; MAX-BW-NEXT: [[TMP16:%.*]] = add nuw nsw i64 [[TMP0]], 0
535
+ ; MAX-BW-NEXT: [[TMP17:%.*]] = add nuw nsw i64 [[TMP1]], 0
536
+ ; MAX-BW-NEXT: [[TMP18:%.*]] = add nuw nsw i64 [[TMP2]], 0
537
+ ; MAX-BW-NEXT: [[TMP19:%.*]] = add nuw nsw i64 [[TMP3]], 0
538
+ ; MAX-BW-NEXT: [[TMP20:%.*]] = add nuw nsw i64 [[TMP4]], 0
539
+ ; MAX-BW-NEXT: [[TMP21:%.*]] = add nuw nsw i64 [[TMP5]], 0
540
+ ; MAX-BW-NEXT: [[TMP22:%.*]] = add nuw nsw i64 [[TMP6]], 0
541
+ ; MAX-BW-NEXT: [[TMP23:%.*]] = add nuw nsw i64 [[TMP7]], 0
542
+ ; MAX-BW-NEXT: [[TMP24:%.*]] = add nuw nsw i64 [[TMP8]], 0
543
+ ; MAX-BW-NEXT: [[TMP25:%.*]] = add nuw nsw i64 [[TMP9]], 0
544
+ ; MAX-BW-NEXT: [[TMP26:%.*]] = add nuw nsw i64 [[TMP10]], 0
545
+ ; MAX-BW-NEXT: [[TMP27:%.*]] = add nuw nsw i64 [[TMP11]], 0
546
+ ; MAX-BW-NEXT: [[TMP28:%.*]] = add nuw nsw i64 [[TMP12]], 0
547
+ ; MAX-BW-NEXT: [[TMP29:%.*]] = add nuw nsw i64 [[TMP13]], 0
548
+ ; MAX-BW-NEXT: [[TMP30:%.*]] = add nuw nsw i64 [[TMP14]], 0
549
+ ; MAX-BW-NEXT: [[TMP31:%.*]] = add nuw nsw i64 [[TMP15]], 0
550
+ ; MAX-BW-NEXT: [[TMP32:%.*]] = getelementptr inbounds [1024 x i32], ptr [[A:%.*]], i64 0, i64 [[TMP16]]
551
+ ; MAX-BW-NEXT: [[TMP33:%.*]] = getelementptr inbounds i32, ptr [[TMP32]], i32 0
552
+ ; MAX-BW-NEXT: [[WIDE_VEC:%.*]] = load <32 x i32>, ptr [[TMP33]], align 4
553
+ ; MAX-BW-NEXT: [[STRIDED_VEC:%.*]] = shufflevector <32 x i32> [[WIDE_VEC]], <32 x i32> poison, <16 x i32> <i32 0, i32 2, i32 4, i32 6, i32 8, i32 10, i32 12, i32 14, i32 16, i32 18, i32 20, i32 22, i32 24, i32 26, i32 28, i32 30>
554
+ ; MAX-BW-NEXT: [[STRIDED_VEC1:%.*]] = shufflevector <32 x i32> [[WIDE_VEC]], <32 x i32> poison, <16 x i32> <i32 1, i32 3, i32 5, i32 7, i32 9, i32 11, i32 13, i32 15, i32 17, i32 19, i32 21, i32 23, i32 25, i32 27, i32 29, i32 31>
555
+ ; MAX-BW-NEXT: [[TMP34:%.*]] = add <16 x i32> [[STRIDED_VEC]], [[STRIDED_VEC1]]
556
+ ; MAX-BW-NEXT: [[TMP35:%.*]] = trunc <16 x i32> [[TMP34]] to <16 x i8>
557
+ ; MAX-BW-NEXT: [[TMP36:%.*]] = getelementptr inbounds [1024 x i8], ptr [[B:%.*]], i64 0, i64 [[TMP16]]
558
+ ; MAX-BW-NEXT: [[TMP37:%.*]] = getelementptr inbounds [1024 x i8], ptr [[B]], i64 0, i64 [[TMP17]]
559
+ ; MAX-BW-NEXT: [[TMP38:%.*]] = getelementptr inbounds [1024 x i8], ptr [[B]], i64 0, i64 [[TMP18]]
560
+ ; MAX-BW-NEXT: [[TMP39:%.*]] = getelementptr inbounds [1024 x i8], ptr [[B]], i64 0, i64 [[TMP19]]
561
+ ; MAX-BW-NEXT: [[TMP40:%.*]] = getelementptr inbounds [1024 x i8], ptr [[B]], i64 0, i64 [[TMP20]]
562
+ ; MAX-BW-NEXT: [[TMP41:%.*]] = getelementptr inbounds [1024 x i8], ptr [[B]], i64 0, i64 [[TMP21]]
563
+ ; MAX-BW-NEXT: [[TMP42:%.*]] = getelementptr inbounds [1024 x i8], ptr [[B]], i64 0, i64 [[TMP22]]
564
+ ; MAX-BW-NEXT: [[TMP43:%.*]] = getelementptr inbounds [1024 x i8], ptr [[B]], i64 0, i64 [[TMP23]]
565
+ ; MAX-BW-NEXT: [[TMP44:%.*]] = getelementptr inbounds [1024 x i8], ptr [[B]], i64 0, i64 [[TMP24]]
566
+ ; MAX-BW-NEXT: [[TMP45:%.*]] = getelementptr inbounds [1024 x i8], ptr [[B]], i64 0, i64 [[TMP25]]
567
+ ; MAX-BW-NEXT: [[TMP46:%.*]] = getelementptr inbounds [1024 x i8], ptr [[B]], i64 0, i64 [[TMP26]]
568
+ ; MAX-BW-NEXT: [[TMP47:%.*]] = getelementptr inbounds [1024 x i8], ptr [[B]], i64 0, i64 [[TMP27]]
569
+ ; MAX-BW-NEXT: [[TMP48:%.*]] = getelementptr inbounds [1024 x i8], ptr [[B]], i64 0, i64 [[TMP28]]
570
+ ; MAX-BW-NEXT: [[TMP49:%.*]] = getelementptr inbounds [1024 x i8], ptr [[B]], i64 0, i64 [[TMP29]]
571
+ ; MAX-BW-NEXT: [[TMP50:%.*]] = getelementptr inbounds [1024 x i8], ptr [[B]], i64 0, i64 [[TMP30]]
572
+ ; MAX-BW-NEXT: [[TMP51:%.*]] = getelementptr inbounds [1024 x i8], ptr [[B]], i64 0, i64 [[TMP31]]
573
+ ; MAX-BW-NEXT: [[TMP52:%.*]] = extractelement <16 x i8> [[TMP35]], i32 0
574
+ ; MAX-BW-NEXT: store i8 [[TMP52]], ptr [[TMP36]], align 1
575
+ ; MAX-BW-NEXT: [[TMP53:%.*]] = extractelement <16 x i8> [[TMP35]], i32 1
576
+ ; MAX-BW-NEXT: store i8 [[TMP53]], ptr [[TMP37]], align 1
577
+ ; MAX-BW-NEXT: [[TMP54:%.*]] = extractelement <16 x i8> [[TMP35]], i32 2
578
+ ; MAX-BW-NEXT: store i8 [[TMP54]], ptr [[TMP38]], align 1
579
+ ; MAX-BW-NEXT: [[TMP55:%.*]] = extractelement <16 x i8> [[TMP35]], i32 3
580
+ ; MAX-BW-NEXT: store i8 [[TMP55]], ptr [[TMP39]], align 1
581
+ ; MAX-BW-NEXT: [[TMP56:%.*]] = extractelement <16 x i8> [[TMP35]], i32 4
582
+ ; MAX-BW-NEXT: store i8 [[TMP56]], ptr [[TMP40]], align 1
583
+ ; MAX-BW-NEXT: [[TMP57:%.*]] = extractelement <16 x i8> [[TMP35]], i32 5
584
+ ; MAX-BW-NEXT: store i8 [[TMP57]], ptr [[TMP41]], align 1
585
+ ; MAX-BW-NEXT: [[TMP58:%.*]] = extractelement <16 x i8> [[TMP35]], i32 6
586
+ ; MAX-BW-NEXT: store i8 [[TMP58]], ptr [[TMP42]], align 1
587
+ ; MAX-BW-NEXT: [[TMP59:%.*]] = extractelement <16 x i8> [[TMP35]], i32 7
588
+ ; MAX-BW-NEXT: store i8 [[TMP59]], ptr [[TMP43]], align 1
589
+ ; MAX-BW-NEXT: [[TMP60:%.*]] = extractelement <16 x i8> [[TMP35]], i32 8
590
+ ; MAX-BW-NEXT: store i8 [[TMP60]], ptr [[TMP44]], align 1
591
+ ; MAX-BW-NEXT: [[TMP61:%.*]] = extractelement <16 x i8> [[TMP35]], i32 9
592
+ ; MAX-BW-NEXT: store i8 [[TMP61]], ptr [[TMP45]], align 1
593
+ ; MAX-BW-NEXT: [[TMP62:%.*]] = extractelement <16 x i8> [[TMP35]], i32 10
594
+ ; MAX-BW-NEXT: store i8 [[TMP62]], ptr [[TMP46]], align 1
595
+ ; MAX-BW-NEXT: [[TMP63:%.*]] = extractelement <16 x i8> [[TMP35]], i32 11
596
+ ; MAX-BW-NEXT: store i8 [[TMP63]], ptr [[TMP47]], align 1
597
+ ; MAX-BW-NEXT: [[TMP64:%.*]] = extractelement <16 x i8> [[TMP35]], i32 12
598
+ ; MAX-BW-NEXT: store i8 [[TMP64]], ptr [[TMP48]], align 1
599
+ ; MAX-BW-NEXT: [[TMP65:%.*]] = extractelement <16 x i8> [[TMP35]], i32 13
600
+ ; MAX-BW-NEXT: store i8 [[TMP65]], ptr [[TMP49]], align 1
601
+ ; MAX-BW-NEXT: [[TMP66:%.*]] = extractelement <16 x i8> [[TMP35]], i32 14
602
+ ; MAX-BW-NEXT: store i8 [[TMP66]], ptr [[TMP50]], align 1
603
+ ; MAX-BW-NEXT: [[TMP67:%.*]] = extractelement <16 x i8> [[TMP35]], i32 15
604
+ ; MAX-BW-NEXT: store i8 [[TMP67]], ptr [[TMP51]], align 1
605
+ ; MAX-BW-NEXT: [[INDEX_NEXT]] = add nuw i64 [[INDEX]], 16
606
+ ; MAX-BW-NEXT: [[TMP68:%.*]] = icmp eq i64 [[INDEX_NEXT]], 512
607
+ ; MAX-BW-NEXT: br i1 [[TMP68]], label [[MIDDLE_BLOCK:%.*]], label [[VECTOR_BODY]], !llvm.loop [[LOOP9:![0-9]+]]
688
608
; MAX-BW: middle.block:
689
609
; MAX-BW-NEXT: br i1 true, label [[FOR_COND_CLEANUP:%.*]], label [[SCALAR_PH]]
690
610
; MAX-BW: scalar.ph:
0 commit comments