[WebAssembly] Fix incorrect assertion in SIMD reduction codegen

[tlively]

Summary

The codegen routine introduced in 18077e9fd688 did not account for vectors with
more than 16 lanes. Remove the incorrect assertion and bail out of the
optimization when encountering this case. Add test cases that previously
triggered the assertion. Unfortunately, these test cases now have terrible
codegen, but that is at least better than crashing.

Fixes #63500.

[tlively]

Issue filed for terrible test codegen: https://github.com/llvm/llvm-project/issues/63606

[aheejin]

Wow the codegen is indeed impressive... 😂

llvm/lib/Target/WebAssembly/WebAssemblyISelLowering.cpp

if (DCI.isBeforeLegalize() && VT.isScalarInteger() &&
       (isNullConstant(RHS) || isAllOnesConstant(RHS)) &&
       LHS->getOpcode() == ISD::BITCAST) {
     EVT FromVT = LHS->getOperand(0).getValueType();
     if (FromVT.isFixedLengthVector() &&
         FromVT.getVectorElementType() == MVT::i1) {
       int Intrin = isNullConstant(RHS) ? Intrinsic::wasm_anytrue
                                        : Intrinsic::wasm_alltrue;
       unsigned NumElts = FromVT.getVectorNumElements();
-      assert(NumElts == 2 || NumElts == 4 || NumElts == 8 || NumElts == 16);
+      if (NumElts != 2 && NumElts != 4 && NumElts != 8 && NumElts != 16)
+        return SDValue();
       EVT Width = MVT::getIntegerVT(128 / NumElts);
       SDValue Ret = DAG.getZExtOrTrunc(
           DAG.getNode(
               ISD::INTRINSIC_WO_CHAIN, DL, MVT::i32,
               {DAG.getConstant(Intrin, DL, MVT::i32),
                DAG.getSExtOrTrunc(LHS->getOperand(0), DL,
                                   FromVT.changeVectorElementType(Width))}),
           DL, MVT::i1);

llvm/test/CodeGen/WebAssembly/simd-vecreduce-bool.ll

 ; NOTE: Assertions have been autogenerated by utils/update_llc_test_checks.py UTC_ARGS: --version 2
 ; RUN: llc < %s -verify-machineinstrs -disable-wasm-fallthrough-return-opt -wasm-disable-explicit-locals -wasm-keep-registers -mattr=+simd128 | FileCheck %s
 
 ; Tests that bool vecreduce produces anytrue and alltrue instructions
 
 target triple = "wasm32-unknown-unknown"
 
 declare i1 @llvm.vector.reduce.or.v2i1(<2 x i1>)
 declare i1 @llvm.vector.reduce.or.v4i1(<4 x i1>)
 declare i1 @llvm.vector.reduce.or.v7i1(<7 x i1>)
 declare i1 @llvm.vector.reduce.or.v8i1(<8 x i1>)
 declare i1 @llvm.vector.reduce.or.v16i1(<16 x i1>)
+declare i1 @llvm.vector.reduce.or.v64i1(<64 x i1>)
 declare i1 @llvm.vector.reduce.and.v2i1(<2 x i1>)
 declare i1 @llvm.vector.reduce.and.v4i1(<4 x i1>)
 declare i1 @llvm.vector.reduce.and.v7i1(<7 x i1>)
 declare i1 @llvm.vector.reduce.and.v8i1(<8 x i1>)
 declare i1 @llvm.vector.reduce.and.v16i1(<16 x i1>)
+declare i1 @llvm.vector.reduce.and.v64i1(<64 x i1> %x)
 
 ; =====================
 ; Regular vectors of i1
 ; =====================
 
 define i1 @test_any_v8i1(<8 x i1> %x) {
 ; CHECK-LABEL: test_any_v8i1:
 ; CHECK:         .functype test_any_v8i1 (v128) -> (i32)

 ; CHECK-NEXT:    i16x8.shr_s $push3=, $pop2, $pop5
 ; CHECK-NEXT:    i16x8.all_true $push4=, $pop3
 ; CHECK-NEXT:    return $pop4
   %bits = trunc <8 x i8> %x to <8 x i1>
   %ret = call i1 @llvm.vector.reduce.and.v8i1(<8 x i1> %bits)
   ret i1 %ret
 }
 
+;; TODO: Improve this codegen.
+define i1 @test_any_v64i8(<64 x i8> %x) {
+; CHECK-LABEL: test_any_v64i8:
+; CHECK:         .functype test_any_v64i8 (v128, v128, v128, v128) -> (i32)
+; CHECK-NEXT:  # %bb.0:
+; CHECK-NEXT:    global.get $push264=, __stack_pointer
+; CHECK-NEXT:    i32.const $push265=, 16
+; CHECK-NEXT:    i32.sub $drop=, $pop264, $pop265
+; CHECK-NEXT:    i8x16.extract_lane_u $push254=, $0, 15
+; CHECK-NEXT:    i8x16.extract_lane_u $push253=, $2, 15
+; CHECK-NEXT:    i32.or $push255=, $pop254, $pop253
+; CHECK-NEXT:    i32.const $push133=, 15
+; CHECK-NEXT:    i32.shl $push256=, $pop255, $pop133
+; CHECK-NEXT:    i8x16.extract_lane_u $push195=, $0, 0
+; CHECK-NEXT:    i32.const $push1=, 1
+; CHECK-NEXT:    i32.and $push196=, $pop195, $pop1
+; CHECK-NEXT:    i8x16.extract_lane_u $push197=, $0, 1
+; CHECK-NEXT:    i32.const $push368=, 1
+; CHECK-NEXT:    i32.and $push198=, $pop197, $pop368
+; CHECK-NEXT:    i32.const $push367=, 1
+; CHECK-NEXT:    i32.shl $push199=, $pop198, $pop367
+; CHECK-NEXT:    i32.or $push200=, $pop196, $pop199
+; CHECK-NEXT:    i8x16.extract_lane_u $push201=, $0, 2
+; CHECK-NEXT:    i32.const $push366=, 1
+; CHECK-NEXT:    i32.and $push202=, $pop201, $pop366
+; CHECK-NEXT:    i32.const $push9=, 2
+; CHECK-NEXT:    i32.shl $push203=, $pop202, $pop9
+; CHECK-NEXT:    i32.or $push204=, $pop200, $pop203
+; CHECK-NEXT:    i8x16.extract_lane_u $push205=, $0, 3
+; CHECK-NEXT:    i32.const $push365=, 1
+; CHECK-NEXT:    i32.and $push206=, $pop205, $pop365
+; CHECK-NEXT:    i32.const $push14=, 3
+; CHECK-NEXT:    i32.shl $push207=, $pop206, $pop14
+; CHECK-NEXT:    i32.or $push208=, $pop204, $pop207
+; CHECK-NEXT:    i8x16.extract_lane_u $push209=, $0, 4
+; CHECK-NEXT:    i32.const $push364=, 1
+; CHECK-NEXT:    i32.and $push210=, $pop209, $pop364
+; CHECK-NEXT:    i32.const $push19=, 4
+; CHECK-NEXT:    i32.shl $push211=, $pop210, $pop19
+; CHECK-NEXT:    i32.or $push212=, $pop208, $pop211
+; CHECK-NEXT:    i8x16.extract_lane_u $push213=, $0, 5
+; CHECK-NEXT:    i32.const $push363=, 1
+; CHECK-NEXT:    i32.and $push214=, $pop213, $pop363
+; CHECK-NEXT:    i32.const $push24=, 5
+; CHECK-NEXT:    i32.shl $push215=, $pop214, $pop24
+; CHECK-NEXT:    i32.or $push216=, $pop212, $pop215
+; CHECK-NEXT:    i8x16.extract_lane_u $push217=, $0, 6
+; CHECK-NEXT:    i32.const $push362=, 1
+; CHECK-NEXT:    i32.and $push218=, $pop217, $pop362
+; CHECK-NEXT:    i32.const $push29=, 6
+; CHECK-NEXT:    i32.shl $push219=, $pop218, $pop29
+; CHECK-NEXT:    i32.or $push220=, $pop216, $pop219
+; CHECK-NEXT:    i8x16.extract_lane_u $push221=, $0, 7
+; CHECK-NEXT:    i32.const $push361=, 1
+; CHECK-NEXT:    i32.and $push222=, $pop221, $pop361
+; CHECK-NEXT:    i32.const $push34=, 7
+; CHECK-NEXT:    i32.shl $push223=, $pop222, $pop34
+; CHECK-NEXT:    i32.or $push224=, $pop220, $pop223
+; CHECK-NEXT:    i8x16.extract_lane_u $push225=, $0, 8
+; CHECK-NEXT:    i32.const $push360=, 1
+; CHECK-NEXT:    i32.and $push226=, $pop225, $pop360
+; CHECK-NEXT:    i32.const $push39=, 8
+; CHECK-NEXT:    i32.shl $push227=, $pop226, $pop39
+; CHECK-NEXT:    i32.or $push228=, $pop224, $pop227
+; CHECK-NEXT:    i8x16.extract_lane_u $push229=, $0, 9
+; CHECK-NEXT:    i32.const $push359=, 1
+; CHECK-NEXT:    i32.and $push230=, $pop229, $pop359
+; CHECK-NEXT:    i32.const $push44=, 9
+; CHECK-NEXT:    i32.shl $push231=, $pop230, $pop44
+; CHECK-NEXT:    i32.or $push232=, $pop228, $pop231
+; CHECK-NEXT:    i8x16.extract_lane_u $push233=, $0, 10
+; CHECK-NEXT:    i32.const $push358=, 1
+; CHECK-NEXT:    i32.and $push234=, $pop233, $pop358
+; CHECK-NEXT:    i32.const $push49=, 10
+; CHECK-NEXT:    i32.shl $push235=, $pop234, $pop49
+; CHECK-NEXT:    i32.or $push236=, $pop232, $pop235
+; CHECK-NEXT:    i8x16.extract_lane_u $push237=, $0, 11
+; CHECK-NEXT:    i32.const $push357=, 1
+; CHECK-NEXT:    i32.and $push238=, $pop237, $pop357
+; CHECK-NEXT:    i32.const $push54=, 11
+; CHECK-NEXT:    i32.shl $push239=, $pop238, $pop54
+; CHECK-NEXT:    i32.or $push240=, $pop236, $pop239
+; CHECK-NEXT:    i8x16.extract_lane_u $push241=, $0, 12
+; CHECK-NEXT:    i32.const $push356=, 1
+; CHECK-NEXT:    i32.and $push242=, $pop241, $pop356
+; CHECK-NEXT:    i32.const $push59=, 12
+; CHECK-NEXT:    i32.shl $push243=, $pop242, $pop59
+; CHECK-NEXT:    i32.or $push244=, $pop240, $pop243
+; CHECK-NEXT:    i8x16.extract_lane_u $push245=, $0, 13
+; CHECK-NEXT:    i32.const $push355=, 1
+; CHECK-NEXT:    i32.and $push246=, $pop245, $pop355
+; CHECK-NEXT:    i32.const $push64=, 13
+; CHECK-NEXT:    i32.shl $push247=, $pop246, $pop64
+; CHECK-NEXT:    i32.or $push248=, $pop244, $pop247
+; CHECK-NEXT:    i8x16.extract_lane_u $push249=, $0, 14
+; CHECK-NEXT:    i32.const $push354=, 1
+; CHECK-NEXT:    i32.and $push250=, $pop249, $pop354
+; CHECK-NEXT:    i32.const $push69=, 14
+; CHECK-NEXT:    i32.shl $push251=, $pop250, $pop69
+; CHECK-NEXT:    i32.or $push252=, $pop248, $pop251
+; CHECK-NEXT:    i32.or $push257=, $pop256, $pop252
+; CHECK-NEXT:    i8x16.extract_lane_u $push137=, $2, 0
+; CHECK-NEXT:    i32.const $push353=, 1
+; CHECK-NEXT:    i32.and $push138=, $pop137, $pop353
+; CHECK-NEXT:    i8x16.extract_lane_u $push139=, $2, 1
+; CHECK-NEXT:    i32.const $push352=, 1
+; CHECK-NEXT:    i32.and $push140=, $pop139, $pop352
+; CHECK-NEXT:    i32.const $push351=, 1
+; CHECK-NEXT:    i32.shl $push141=, $pop140, $pop351
+; CHECK-NEXT:    i32.or $push142=, $pop138, $pop141
+; CHECK-NEXT:    i8x16.extract_lane_u $push143=, $2, 2
+; CHECK-NEXT:    i32.const $push350=, 1
+; CHECK-NEXT:    i32.and $push144=, $pop143, $pop350
+; CHECK-NEXT:    i32.const $push349=, 2
+; CHECK-NEXT:    i32.shl $push145=, $pop144, $pop349
+; CHECK-NEXT:    i32.or $push146=, $pop142, $pop145
+; CHECK-NEXT:    i8x16.extract_lane_u $push147=, $2, 3
+; CHECK-NEXT:    i32.const $push348=, 1
+; CHECK-NEXT:    i32.and $push148=, $pop147, $pop348
+; CHECK-NEXT:    i32.const $push347=, 3
+; CHECK-NEXT:    i32.shl $push149=, $pop148, $pop347
+; CHECK-NEXT:    i32.or $push150=, $pop146, $pop149
+; CHECK-NEXT:    i8x16.extract_lane_u $push151=, $2, 4
+; CHECK-NEXT:    i32.const $push346=, 1
+; CHECK-NEXT:    i32.and $push152=, $pop151, $pop346
+; CHECK-NEXT:    i32.const $push345=, 4
+; CHECK-NEXT:    i32.shl $push153=, $pop152, $pop345
+; CHECK-NEXT:    i32.or $push154=, $pop150, $pop153
+; CHECK-NEXT:    i8x16.extract_lane_u $push155=, $2, 5
+; CHECK-NEXT:    i32.const $push344=, 1
+; CHECK-NEXT:    i32.and $push156=, $pop155, $pop344
+; CHECK-NEXT:    i32.const $push343=, 5
+; CHECK-NEXT:    i32.shl $push157=, $pop156, $pop343
+; CHECK-NEXT:    i32.or $push158=, $pop154, $pop157
+; CHECK-NEXT:    i8x16.extract_lane_u $push159=, $2, 6
+; CHECK-NEXT:    i32.const $push342=, 1
+; CHECK-NEXT:    i32.and $push160=, $pop159, $pop342
+; CHECK-NEXT:    i32.const $push341=, 6
+; CHECK-NEXT:    i32.shl $push161=, $pop160, $pop341
+; CHECK-NEXT:    i32.or $push162=, $pop158, $pop161
+; CHECK-NEXT:    i8x16.extract_lane_u $push163=, $2, 7
+; CHECK-NEXT:    i32.const $push340=, 1
+; CHECK-NEXT:    i32.and $push164=, $pop163, $pop340
+; CHECK-NEXT:    i32.const $push339=, 7
+; CHECK-NEXT:    i32.shl $push165=, $pop164, $pop339
+; CHECK-NEXT:    i32.or $push166=, $pop162, $pop165
+; CHECK-NEXT:    i8x16.extract_lane_u $push167=, $2, 8
+; CHECK-NEXT:    i32.const $push338=, 1
+; CHECK-NEXT:    i32.and $push168=, $pop167, $pop338
+; CHECK-NEXT:    i32.const $push337=, 8
+; CHECK-NEXT:    i32.shl $push169=, $pop168, $pop337
+; CHECK-NEXT:    i32.or $push170=, $pop166, $pop169
+; CHECK-NEXT:    i8x16.extract_lane_u $push171=, $2, 9
+; CHECK-NEXT:    i32.const $push336=, 1
+; CHECK-NEXT:    i32.and $push172=, $pop171, $pop336
+; CHECK-NEXT:    i32.const $push335=, 9
+; CHECK-NEXT:    i32.shl $push173=, $pop172, $pop335
+; CHECK-NEXT:    i32.or $push174=, $pop170, $pop173
+; CHECK-NEXT:    i8x16.extract_lane_u $push175=, $2, 10
+; CHECK-NEXT:    i32.const $push334=, 1
+; CHECK-NEXT:    i32.and $push176=, $pop175, $pop334
+; CHECK-NEXT:    i32.const $push333=, 10
+; CHECK-NEXT:    i32.shl $push177=, $pop176, $pop333
+; CHECK-NEXT:    i32.or $push178=, $pop174, $pop177
+; CHECK-NEXT:    i8x16.extract_lane_u $push179=, $2, 11
+; CHECK-NEXT:    i32.const $push332=, 1
+; CHECK-NEXT:    i32.and $push180=, $pop179, $pop332
+; CHECK-NEXT:    i32.const $push331=, 11
+; CHECK-NEXT:    i32.shl $push181=, $pop180, $pop331
+; CHECK-NEXT:    i32.or $push182=, $pop178, $pop181
+; CHECK-NEXT:    i8x16.extract_lane_u $push183=, $2, 12
+; CHECK-NEXT:    i32.const $push330=, 1
+; CHECK-NEXT:    i32.and $push184=, $pop183, $pop330
+; CHECK-NEXT:    i32.const $push329=, 12
+; CHECK-NEXT:    i32.shl $push185=, $pop184, $pop329
+; CHECK-NEXT:    i32.or $push186=, $pop182, $pop185
+; CHECK-NEXT:    i8x16.extract_lane_u $push187=, $2, 13
+; CHECK-NEXT:    i32.const $push328=, 1
+; CHECK-NEXT:    i32.and $push188=, $pop187, $pop328
+; CHECK-NEXT:    i32.const $push327=, 13
+; CHECK-NEXT:    i32.shl $push189=, $pop188, $pop327
+; CHECK-NEXT:    i32.or $push190=, $pop186, $pop189
+; CHECK-NEXT:    i8x16.extract_lane_u $push191=, $2, 14
+; CHECK-NEXT:    i32.const $push326=, 1
+; CHECK-NEXT:    i32.and $push192=, $pop191, $pop326
+; CHECK-NEXT:    i32.const $push325=, 14
+; CHECK-NEXT:    i32.shl $push193=, $pop192, $pop325
+; CHECK-NEXT:    i32.or $push194=, $pop190, $pop193
+; CHECK-NEXT:    i32.or $push258=, $pop257, $pop194
+; CHECK-NEXT:    i8x16.extract_lane_u $push131=, $1, 15
+; CHECK-NEXT:    i8x16.extract_lane_u $push130=, $3, 15
+; CHECK-NEXT:    i32.or $push132=, $pop131, $pop130
+; CHECK-NEXT:    i32.const $push324=, 15
+; CHECK-NEXT:    i32.shl $push134=, $pop132, $pop324
+; CHECK-NEXT:    i8x16.extract_lane_u $push72=, $1, 0
+; CHECK-NEXT:    i32.const $push323=, 1
+; CHECK-NEXT:    i32.and $push73=, $pop72, $pop323
+; CHECK-NEXT:    i8x16.extract_lane_u $push74=, $1, 1
+; CHECK-NEXT:    i32.const $push322=, 1
+; CHECK-NEXT:    i32.and $push75=, $pop74, $pop322
+; CHECK-NEXT:    i32.const $push321=, 1
+; CHECK-NEXT:    i32.shl $push76=, $pop75, $pop321
+; CHECK-NEXT:    i32.or $push77=, $pop73, $pop76
+; CHECK-NEXT:    i8x16.extract_lane_u $push78=, $1, 2
+; CHECK-NEXT:    i32.const $push320=, 1
+; CHECK-NEXT:    i32.and $push79=, $pop78, $pop320
+; CHECK-NEXT:    i32.const $push319=, 2
+; CHECK-NEXT:    i32.shl $push80=, $pop79, $pop319
+; CHECK-NEXT:    i32.or $push81=, $pop77, $pop80
+; CHECK-NEXT:    i8x16.extract_lane_u $push82=, $1, 3
+; CHECK-NEXT:    i32.const $push318=, 1
+; CHECK-NEXT:    i32.and $push83=, $pop82, $pop318
+; CHECK-NEXT:    i32.const $push317=, 3
+; CHECK-NEXT:    i32.shl $push84=, $pop83, $pop317
+; CHECK-NEXT:    i32.or $push85=, $pop81, $pop84
+; CHECK-NEXT:    i8x16.extract_lane_u $push86=, $1, 4
+; CHECK-NEXT:    i32.const $push316=, 1
+; CHECK-NEXT:    i32.and $push87=, $pop86, $pop316
+; CHECK-NEXT:    i32.const $push315=, 4
+; CHECK-NEXT:    i32.shl $push88=, $pop87, $pop315
+; CHECK-NEXT:    i32.or $push89=, $pop85, $pop88
+; CHECK-NEXT:    i8x16.extract_lane_u $push90=, $1, 5
+; CHECK-NEXT:    i32.const $push314=, 1
+; CHECK-NEXT:    i32.and $push91=, $pop90, $pop314
+; CHECK-NEXT:    i32.const $push313=, 5
+; CHECK-NEXT:    i32.shl $push92=, $pop91, $pop313
+; CHECK-NEXT:    i32.or $push93=, $pop89, $pop92
+; CHECK-NEXT:    i8x16.extract_lane_u $push94=, $1, 6
+; CHECK-NEXT:    i32.const $push312=, 1
+; CHECK-NEXT:    i32.and $push95=, $pop94, $pop312
+; CHECK-NEXT:    i32.const $push311=, 6
+; CHECK-NEXT:    i32.shl $push96=, $pop95, $pop311
+; CHECK-NEXT:    i32.or $push97=, $pop93, $pop96
+; CHECK-NEXT:    i8x16.extract_lane_u $push98=, $1, 7
+; CHECK-NEXT:    i32.const $push310=, 1
+; CHECK-NEXT:    i32.and $push99=, $pop98, $pop310
+; CHECK-NEXT:    i32.const $push309=, 7
+; CHECK-NEXT:    i32.shl $push100=, $pop99, $pop309
+; CHECK-NEXT:    i32.or $push101=, $pop97, $pop100
+; CHECK-NEXT:    i8x16.extract_lane_u $push102=, $1, 8
+; CHECK-NEXT:    i32.const $push308=, 1
+; CHECK-NEXT:    i32.and $push103=, $pop102, $pop308
+; CHECK-NEXT:    i32.const $push307=, 8
+; CHECK-NEXT:    i32.shl $push104=, $pop103, $pop307
+; CHECK-NEXT:    i32.or $push105=, $pop101, $pop104
+; CHECK-NEXT:    i8x16.extract_lane_u $push106=, $1, 9
+; CHECK-NEXT:    i32.const $push306=, 1
+; CHECK-NEXT:    i32.and $push107=, $pop106, $pop306
+; CHECK-NEXT:    i32.const $push305=, 9
+; CHECK-NEXT:    i32.shl $push108=, $pop107, $pop305
+; CHECK-NEXT:    i32.or $push109=, $pop105, $pop108
+; CHECK-NEXT:    i8x16.extract_lane_u $push110=, $1, 10
+; CHECK-NEXT:    i32.const $push304=, 1
+; CHECK-NEXT:    i32.and $push111=, $pop110, $pop304
+; CHECK-NEXT:    i32.const $push303=, 10
+; CHECK-NEXT:    i32.shl $push112=, $pop111, $pop303
+; CHECK-NEXT:    i32.or $push113=, $pop109, $pop112
+; CHECK-NEXT:    i8x16.extract_lane_u $push114=, $1, 11
+; CHECK-NEXT:    i32.const $push302=, 1
+; CHECK-NEXT:    i32.and $push115=, $pop114, $pop302
+; CHECK-NEXT:    i32.const $push301=, 11
+; CHECK-NEXT:    i32.shl $push116=, $pop115, $pop301
+; CHECK-NEXT:    i32.or $push117=, $pop113, $pop116
+; CHECK-NEXT:    i8x16.extract_lane_u $push118=, $1, 12
+; CHECK-NEXT:    i32.const $push300=, 1
+; CHECK-NEXT:    i32.and $push119=, $pop118, $pop300
+; CHECK-NEXT:    i32.const $push299=, 12
+; CHECK-NEXT:    i32.shl $push120=, $pop119, $pop299
+; CHECK-NEXT:    i32.or $push121=, $pop117, $pop120
+; CHECK-NEXT:    i8x16.extract_lane_u $push122=, $1, 13
+; CHECK-NEXT:    i32.const $push298=, 1
+; CHECK-NEXT:    i32.and $push123=, $pop122, $pop298
+; CHECK-NEXT:    i32.const $push297=, 13
+; CHECK-NEXT:    i32.shl $push124=, $pop123, $pop297
+; CHECK-NEXT:    i32.or $push125=, $pop121, $pop124
+; CHECK-NEXT:    i8x16.extract_lane_u $push126=, $1, 14
+; CHECK-NEXT:    i32.const $push296=, 1
+; CHECK-NEXT:    i32.and $push127=, $pop126, $pop296
+; CHECK-NEXT:    i32.const $push295=, 14
+; CHECK-NEXT:    i32.shl $push128=, $pop127, $pop295
+; CHECK-NEXT:    i32.or $push129=, $pop125, $pop128
+; CHECK-NEXT:    i32.or $push135=, $pop134, $pop129
+; CHECK-NEXT:    i8x16.extract_lane_u $push0=, $3, 0
+; CHECK-NEXT:    i32.const $push294=, 1
+; CHECK-NEXT:    i32.and $push2=, $pop0, $pop294
+; CHECK-NEXT:    i8x16.extract_lane_u $push3=, $3, 1
+; CHECK-NEXT:    i32.const $push293=, 1
+; CHECK-NEXT:    i32.and $push4=, $pop3, $pop293
+; CHECK-NEXT:    i32.const $push292=, 1
+; CHECK-NEXT:    i32.shl $push5=, $pop4, $pop292
+; CHECK-NEXT:    i32.or $push6=, $pop2, $pop5
+; CHECK-NEXT:    i8x16.extract_lane_u $push7=, $3, 2
+; CHECK-NEXT:    i32.const $push291=, 1
+; CHECK-NEXT:    i32.and $push8=, $pop7, $pop291
+; CHECK-NEXT:    i32.const $push290=, 2
+; CHECK-NEXT:    i32.shl $push10=, $pop8, $pop290
+; CHECK-NEXT:    i32.or $push11=, $pop6, $pop10
+; CHECK-NEXT:    i8x16.extract_lane_u $push12=, $3, 3
+; CHECK-NEXT:    i32.const $push289=, 1
+; CHECK-NEXT:    i32.and $push13=, $pop12, $pop289
+; CHECK-NEXT:    i32.const $push288=, 3
+; CHECK-NEXT:    i32.shl $push15=, $pop13, $pop288
+; CHECK-NEXT:    i32.or $push16=, $pop11, $pop15
+; CHECK-NEXT:    i8x16.extract_lane_u $push17=, $3, 4
+; CHECK-NEXT:    i32.const $push287=, 1
+; CHECK-NEXT:    i32.and $push18=, $pop17, $pop287
+; CHECK-NEXT:    i32.const $push286=, 4
+; CHECK-NEXT:    i32.shl $push20=, $pop18, $pop286
+; CHECK-NEXT:    i32.or $push21=, $pop16, $pop20
+; CHECK-NEXT:    i8x16.extract_lane_u $push22=, $3, 5
+; CHECK-NEXT:    i32.const $push285=, 1
+; CHECK-NEXT:    i32.and $push23=, $pop22, $pop285
+; CHECK-NEXT:    i32.const $push284=, 5
+; CHECK-NEXT:    i32.shl $push25=, $pop23, $pop284
+; CHECK-NEXT:    i32.or $push26=, $pop21, $pop25
+; CHECK-NEXT:    i8x16.extract_lane_u $push27=, $3, 6
+; CHECK-NEXT:    i32.const $push283=, 1
+; CHECK-NEXT:    i32.and $push28=, $pop27, $pop283
+; CHECK-NEXT:    i32.const $push282=, 6
+; CHECK-NEXT:    i32.shl $push30=, $pop28, $pop282
+; CHECK-NEXT:    i32.or $push31=, $pop26, $pop30
+; CHECK-NEXT:    i8x16.extract_lane_u $push32=, $3, 7
+; CHECK-NEXT:    i32.const $push281=, 1
+; CHECK-NEXT:    i32.and $push33=, $pop32, $pop281
+; CHECK-NEXT:    i32.const $push280=, 7
+; CHECK-NEXT:    i32.shl $push35=, $pop33, $pop280
+; CHECK-NEXT:    i32.or $push36=, $pop31, $pop35
+; CHECK-NEXT:    i8x16.extract_lane_u $push37=, $3, 8
+; CHECK-NEXT:    i32.const $push279=, 1
+; CHECK-NEXT:    i32.and $push38=, $pop37, $pop279
+; CHECK-NEXT:    i32.const $push278=, 8
+; CHECK-NEXT:    i32.shl $push40=, $pop38, $pop278
+; CHECK-NEXT:    i32.or $push41=, $pop36, $pop40
+; CHECK-NEXT:    i8x16.extract_lane_u $push42=, $3, 9
+; CHECK-NEXT:    i32.const $push277=, 1
+; CHECK-NEXT:    i32.and $push43=, $pop42, $pop277
+; CHECK-NEXT:    i32.const $push276=, 9
+; CHECK-NEXT:    i32.shl $push45=, $pop43, $pop276
+; CHECK-NEXT:    i32.or $push46=, $pop41, $pop45
+; CHECK-NEXT:    i8x16.extract_lane_u $push47=, $3, 10
+; CHECK-NEXT:    i32.const $push275=, 1
+; CHECK-NEXT:    i32.and $push48=, $pop47, $pop275
+; CHECK-NEXT:    i32.const $push274=, 10
+; CHECK-NEXT:    i32.shl $push50=, $pop48, $pop274
+; CHECK-NEXT:    i32.or $push51=, $pop46, $pop50
+; CHECK-NEXT:    i8x16.extract_lane_u $push52=, $3, 11
+; CHECK-NEXT:    i32.const $push273=, 1
+; CHECK-NEXT:    i32.and $push53=, $pop52, $pop273
+; CHECK-NEXT:    i32.const $push272=, 11
+; CHECK-NEXT:    i32.shl $push55=, $pop53, $pop272
+; CHECK-NEXT:    i32.or $push56=, $pop51, $pop55
+; CHECK-NEXT:    i8x16.extract_lane_u $push57=, $3, 12
+; CHECK-NEXT:    i32.const $push271=, 1
+; CHECK-NEXT:    i32.and $push58=, $pop57, $pop271
+; CHECK-NEXT:    i32.const $push270=, 12
+; CHECK-NEXT:    i32.shl $push60=, $pop58, $pop270
+; CHECK-NEXT:    i32.or $push61=, $pop56, $pop60
+; CHECK-NEXT:    i8x16.extract_lane_u $push62=, $3, 13
+; CHECK-NEXT:    i32.const $push269=, 1
+; CHECK-NEXT:    i32.and $push63=, $pop62, $pop269
+; CHECK-NEXT:    i32.const $push268=, 13
+; CHECK-NEXT:    i32.shl $push65=, $pop63, $pop268
+; CHECK-NEXT:    i32.or $push66=, $pop61, $pop65
+; CHECK-NEXT:    i8x16.extract_lane_u $push67=, $3, 14
+; CHECK-NEXT:    i32.const $push267=, 1
+; CHECK-NEXT:    i32.and $push68=, $pop67, $pop267
+; CHECK-NEXT:    i32.const $push266=, 14
+; CHECK-NEXT:    i32.shl $push70=, $pop68, $pop266
+; CHECK-NEXT:    i32.or $push71=, $pop66, $pop70
+; CHECK-NEXT:    i32.or $push136=, $pop135, $pop71
+; CHECK-NEXT:    i32.or $push259=, $pop258, $pop136
+; CHECK-NEXT:    i32.const $push260=, 65535
+; CHECK-NEXT:    i32.and $push261=, $pop259, $pop260
+; CHECK-NEXT:    i32.const $push262=, 0
+; CHECK-NEXT:    i32.ne $push263=, $pop261, $pop262
+; CHECK-NEXT:    return $pop263
+  %bits = trunc <64 x i8> %x to <64 x i1>
+  %ret = call i1 @llvm.vector.reduce.or.v64i1(<64 x i1> %bits)
+  ret i1 %ret
+}
+
+;; TODO: Improve this codegen.
+define i1 @test_all_v64i8(<64 x i8> %x) {
+; CHECK-LABEL: test_all_v64i8:
+; CHECK:         .functype test_all_v64i8 (v128, v128, v128, v128) -> (i32)
+; CHECK-NEXT:  # %bb.0:
+; CHECK-NEXT:    global.get $push287=, __stack_pointer
+; CHECK-NEXT:    i32.const $push288=, 16
+; CHECK-NEXT:    i32.sub $drop=, $pop287, $pop288
+; CHECK-NEXT:    i8x16.extract_lane_u $push220=, $2, 0
+; CHECK-NEXT:    i32.const $push1=, 1
+; CHECK-NEXT:    i32.and $push221=, $pop220, $pop1
+; CHECK-NEXT:    i8x16.extract_lane_u $push222=, $2, 1
+; CHECK-NEXT:    i32.const $push380=, 1
+; CHECK-NEXT:    i32.and $push223=, $pop222, $pop380
+; CHECK-NEXT:    i32.const $push379=, 1
+; CHECK-NEXT:    i32.shl $push224=, $pop223, $pop379
+; CHECK-NEXT:    i32.or $push225=, $pop221, $pop224
+; CHECK-NEXT:    i8x16.extract_lane_u $push226=, $2, 2
+; CHECK-NEXT:    i32.const $push378=, 1
+; CHECK-NEXT:    i32.and $push227=, $pop226, $pop378
+; CHECK-NEXT:    i32.const $push87=, 2
+; CHECK-NEXT:    i32.shl $push228=, $pop227, $pop87
+; CHECK-NEXT:    i32.or $push229=, $pop225, $pop228
+; CHECK-NEXT:    i8x16.extract_lane_u $push230=, $2, 3
+; CHECK-NEXT:    i32.const $push377=, 1
+; CHECK-NEXT:    i32.and $push231=, $pop230, $pop377
+; CHECK-NEXT:    i32.const $push92=, 3
+; CHECK-NEXT:    i32.shl $push232=, $pop231, $pop92
+; CHECK-NEXT:    i32.or $push233=, $pop229, $pop232
+; CHECK-NEXT:    i8x16.extract_lane_u $push234=, $2, 4
+; CHECK-NEXT:    i32.const $push376=, 1
+; CHECK-NEXT:    i32.and $push235=, $pop234, $pop376
+; CHECK-NEXT:    i32.const $push97=, 4
+; CHECK-NEXT:    i32.shl $push236=, $pop235, $pop97
+; CHECK-NEXT:    i32.or $push237=, $pop233, $pop236
+; CHECK-NEXT:    i8x16.extract_lane_u $push238=, $2, 5
+; CHECK-NEXT:    i32.const $push375=, 1
+; CHECK-NEXT:    i32.and $push239=, $pop238, $pop375
+; CHECK-NEXT:    i32.const $push102=, 5
+; CHECK-NEXT:    i32.shl $push240=, $pop239, $pop102
+; CHECK-NEXT:    i32.or $push241=, $pop237, $pop240
+; CHECK-NEXT:    i8x16.extract_lane_u $push242=, $2, 6
+; CHECK-NEXT:    i32.const $push374=, 1
+; CHECK-NEXT:    i32.and $push243=, $pop242, $pop374
+; CHECK-NEXT:    i32.const $push107=, 6
+; CHECK-NEXT:    i32.shl $push244=, $pop243, $pop107
+; CHECK-NEXT:    i32.or $push245=, $pop241, $pop244
+; CHECK-NEXT:    i8x16.extract_lane_u $push246=, $2, 7
+; CHECK-NEXT:    i32.const $push373=, 1
+; CHECK-NEXT:    i32.and $push247=, $pop246, $pop373
+; CHECK-NEXT:    i32.const $push112=, 7
+; CHECK-NEXT:    i32.shl $push248=, $pop247, $pop112
+; CHECK-NEXT:    i32.or $push249=, $pop245, $pop248
+; CHECK-NEXT:    i8x16.extract_lane_u $push250=, $2, 8
+; CHECK-NEXT:    i32.const $push372=, 1
+; CHECK-NEXT:    i32.and $push251=, $pop250, $pop372
+; CHECK-NEXT:    i32.const $push117=, 8
+; CHECK-NEXT:    i32.shl $push252=, $pop251, $pop117
+; CHECK-NEXT:    i32.or $push253=, $pop249, $pop252
+; CHECK-NEXT:    i8x16.extract_lane_u $push254=, $2, 9
+; CHECK-NEXT:    i32.const $push371=, 1
+; CHECK-NEXT:    i32.and $push255=, $pop254, $pop371
+; CHECK-NEXT:    i32.const $push122=, 9
+; CHECK-NEXT:    i32.shl $push256=, $pop255, $pop122
+; CHECK-NEXT:    i32.or $push257=, $pop253, $pop256
+; CHECK-NEXT:    i8x16.extract_lane_u $push258=, $2, 10
+; CHECK-NEXT:    i32.const $push370=, 1
+; CHECK-NEXT:    i32.and $push259=, $pop258, $pop370
+; CHECK-NEXT:    i32.const $push127=, 10
+; CHECK-NEXT:    i32.shl $push260=, $pop259, $pop127
+; CHECK-NEXT:    i32.or $push261=, $pop257, $pop260
+; CHECK-NEXT:    i8x16.extract_lane_u $push262=, $2, 11
+; CHECK-NEXT:    i32.const $push369=, 1
+; CHECK-NEXT:    i32.and $push263=, $pop262, $pop369
+; CHECK-NEXT:    i32.const $push132=, 11
+; CHECK-NEXT:    i32.shl $push264=, $pop263, $pop132
+; CHECK-NEXT:    i32.or $push265=, $pop261, $pop264
+; CHECK-NEXT:    i8x16.extract_lane_u $push266=, $2, 12
+; CHECK-NEXT:    i32.const $push368=, 1
+; CHECK-NEXT:    i32.and $push267=, $pop266, $pop368
+; CHECK-NEXT:    i32.const $push137=, 12
+; CHECK-NEXT:    i32.shl $push268=, $pop267, $pop137
+; CHECK-NEXT:    i32.or $push269=, $pop265, $pop268
+; CHECK-NEXT:    i8x16.extract_lane_u $push270=, $2, 13
+; CHECK-NEXT:    i32.const $push367=, 1
+; CHECK-NEXT:    i32.and $push271=, $pop270, $pop367
+; CHECK-NEXT:    i32.const $push142=, 13
+; CHECK-NEXT:    i32.shl $push272=, $pop271, $pop142
+; CHECK-NEXT:    i32.or $push273=, $pop269, $pop272
+; CHECK-NEXT:    i8x16.extract_lane_u $push274=, $2, 14
+; CHECK-NEXT:    i32.const $push366=, 1
+; CHECK-NEXT:    i32.and $push275=, $pop274, $pop366
+; CHECK-NEXT:    i32.const $push147=, 14
+; CHECK-NEXT:    i32.shl $push276=, $pop275, $pop147
+; CHECK-NEXT:    i32.or $push277=, $pop273, $pop276
+; CHECK-NEXT:    i8x16.extract_lane_u $push278=, $2, 15
+; CHECK-NEXT:    i32.const $push151=, 15
+; CHECK-NEXT:    i32.shl $push279=, $pop278, $pop151
+; CHECK-NEXT:    i32.or $push280=, $pop277, $pop279
+; CHECK-NEXT:    i32.const $push154=, 65535
+; CHECK-NEXT:    i32.and $push281=, $pop280, $pop154
+; CHECK-NEXT:    i8x16.extract_lane_u $push217=, $3, 15
+; CHECK-NEXT:    i32.const $push76=, 31
+; CHECK-NEXT:    i32.shl $push218=, $pop217, $pop76
+; CHECK-NEXT:    i8x16.extract_lane_u $push213=, $3, 14
+; CHECK-NEXT:    i32.const $push365=, 1
+; CHECK-NEXT:    i32.and $push214=, $pop213, $pop365
+; CHECK-NEXT:    i32.const $push72=, 30
+; CHECK-NEXT:    i32.shl $push215=, $pop214, $pop72
+; CHECK-NEXT:    i8x16.extract_lane_u $push209=, $3, 13
+; CHECK-NEXT:    i32.const $push364=, 1
+; CHECK-NEXT:    i32.and $push210=, $pop209, $pop364
+; CHECK-NEXT:    i32.const $push67=, 29
+; CHECK-NEXT:    i32.shl $push211=, $pop210, $pop67
+; CHECK-NEXT:    i8x16.extract_lane_u $push205=, $3, 12
+; CHECK-NEXT:    i32.const $push363=, 1
+; CHECK-NEXT:    i32.and $push206=, $pop205, $pop363
+; CHECK-NEXT:    i32.const $push62=, 28
+; CHECK-NEXT:    i32.shl $push207=, $pop206, $pop62
+; CHECK-NEXT:    i8x16.extract_lane_u $push201=, $3, 11
+; CHECK-NEXT:    i32.const $push362=, 1
+; CHECK-NEXT:    i32.and $push202=, $pop201, $pop362
+; CHECK-NEXT:    i32.const $push57=, 27
+; CHECK-NEXT:    i32.shl $push203=, $pop202, $pop57
+; CHECK-NEXT:    i8x16.extract_lane_u $push197=, $3, 10
+; CHECK-NEXT:    i32.const $push361=, 1
+; CHECK-NEXT:    i32.and $push198=, $pop197, $pop361
+; CHECK-NEXT:    i32.const $push52=, 26
+; CHECK-NEXT:    i32.shl $push199=, $pop198, $pop52
+; CHECK-NEXT:    i8x16.extract_lane_u $push193=, $3, 9
+; CHECK-NEXT:    i32.const $push360=, 1
+; CHECK-NEXT:    i32.and $push194=, $pop193, $pop360
+; CHECK-NEXT:    i32.const $push47=, 25
+; CHECK-NEXT:    i32.shl $push195=, $pop194, $pop47
+; CHECK-NEXT:    i8x16.extract_lane_u $push189=, $3, 8
+; CHECK-NEXT:    i32.const $push359=, 1
+; CHECK-NEXT:    i32.and $push190=, $pop189, $pop359
+; CHECK-NEXT:    i32.const $push42=, 24
+; CHECK-NEXT:    i32.shl $push191=, $pop190, $pop42
+; CHECK-NEXT:    i8x16.extract_lane_u $push185=, $3, 7
+; CHECK-NEXT:    i32.const $push358=, 1
+; CHECK-NEXT:    i32.and $push186=, $pop185, $pop358
+; CHECK-NEXT:    i32.const $push37=, 23
+; CHECK-NEXT:    i32.shl $push187=, $pop186, $pop37
+; CHECK-NEXT:    i8x16.extract_lane_u $push181=, $3, 6
+; CHECK-NEXT:    i32.const $push357=, 1
+; CHECK-NEXT:    i32.and $push182=, $pop181, $pop357
+; CHECK-NEXT:    i32.const $push32=, 22
+; CHECK-NEXT:    i32.shl $push183=, $pop182, $pop32
+; CHECK-NEXT:    i8x16.extract_lane_u $push177=, $3, 5
+; CHECK-NEXT:    i32.const $push356=, 1
+; CHECK-NEXT:    i32.and $push178=, $pop177, $pop356
+; CHECK-NEXT:    i32.const $push27=, 21
+; CHECK-NEXT:    i32.shl $push179=, $pop178, $pop27
+; CHECK-NEXT:    i8x16.extract_lane_u $push173=, $3, 4
+; CHECK-NEXT:    i32.const $push355=, 1
+; CHECK-NEXT:    i32.and $push174=, $pop173, $pop355
+; CHECK-NEXT:    i32.const $push22=, 20
+; CHECK-NEXT:    i32.shl $push175=, $pop174, $pop22
+; CHECK-NEXT:    i8x16.extract_lane_u $push169=, $3, 3
+; CHECK-NEXT:    i32.const $push354=, 1
+; CHECK-NEXT:    i32.and $push170=, $pop169, $pop354
+; CHECK-NEXT:    i32.const $push17=, 19
+; CHECK-NEXT:    i32.shl $push171=, $pop170, $pop17
+; CHECK-NEXT:    i8x16.extract_lane_u $push165=, $3, 2
+; CHECK-NEXT:    i32.const $push353=, 1
+; CHECK-NEXT:    i32.and $push166=, $pop165, $pop353
+; CHECK-NEXT:    i32.const $push12=, 18
+; CHECK-NEXT:    i32.shl $push167=, $pop166, $pop12
+; CHECK-NEXT:    i8x16.extract_lane_u $push161=, $3, 1
+; CHECK-NEXT:    i32.const $push352=, 1
+; CHECK-NEXT:    i32.and $push162=, $pop161, $pop352
+; CHECK-NEXT:    i32.const $push7=, 17
+; CHECK-NEXT:    i32.shl $push163=, $pop162, $pop7
+; CHECK-NEXT:    i8x16.extract_lane_u $push158=, $3, 0
+; CHECK-NEXT:    i32.const $push351=, 1
+; CHECK-NEXT:    i32.and $push159=, $pop158, $pop351
+; CHECK-NEXT:    i32.const $push3=, 16
+; CHECK-NEXT:    i32.shl $push160=, $pop159, $pop3
+; CHECK-NEXT:    i32.or $push164=, $pop163, $pop160
+; CHECK-NEXT:    i32.or $push168=, $pop167, $pop164
+; CHECK-NEXT:    i32.or $push172=, $pop171, $pop168
+; CHECK-NEXT:    i32.or $push176=, $pop175, $pop172
+; CHECK-NEXT:    i32.or $push180=, $pop179, $pop176
+; CHECK-NEXT:    i32.or $push184=, $pop183, $pop180
+; CHECK-NEXT:    i32.or $push188=, $pop187, $pop184
+; CHECK-NEXT:    i32.or $push192=, $pop191, $pop188
+; CHECK-NEXT:    i32.or $push196=, $pop195, $pop192
+; CHECK-NEXT:    i32.or $push200=, $pop199, $pop196
+; CHECK-NEXT:    i32.or $push204=, $pop203, $pop200
+; CHECK-NEXT:    i32.or $push208=, $pop207, $pop204
+; CHECK-NEXT:    i32.or $push212=, $pop211, $pop208
+; CHECK-NEXT:    i32.or $push216=, $pop215, $pop212
+; CHECK-NEXT:    i32.or $push219=, $pop218, $pop216
+; CHECK-NEXT:    i32.or $push282=, $pop281, $pop219
+; CHECK-NEXT:    i64.extend_i32_u $push283=, $pop282
+; CHECK-NEXT:    i8x16.extract_lane_u $push79=, $0, 0
+; CHECK-NEXT:    i32.const $push350=, 1
+; CHECK-NEXT:    i32.and $push80=, $pop79, $pop350
+; CHECK-NEXT:    i8x16.extract_lane_u $push81=, $0, 1
+; CHECK-NEXT:    i32.const $push349=, 1
+; CHECK-NEXT:    i32.and $push82=, $pop81, $pop349
+; CHECK-NEXT:    i32.const $push348=, 1
+; CHECK-NEXT:    i32.shl $push83=, $pop82, $pop348
+; CHECK-NEXT:    i32.or $push84=, $pop80, $pop83
+; CHECK-NEXT:    i8x16.extract_lane_u $push85=, $0, 2
+; CHECK-NEXT:    i32.const $push347=, 1
+; CHECK-NEXT:    i32.and $push86=, $pop85, $pop347
+; CHECK-NEXT:    i32.const $push346=, 2
+; CHECK-NEXT:    i32.shl $push88=, $pop86, $pop346
+; CHECK-NEXT:    i32.or $push89=, $pop84, $pop88
+; CHECK-NEXT:    i8x16.extract_lane_u $push90=, $0, 3
+; CHECK-NEXT:    i32.const $push345=, 1
+; CHECK-NEXT:    i32.and $push91=, $pop90, $pop345
+; CHECK-NEXT:    i32.const $push344=, 3
+; CHECK-NEXT:    i32.shl $push93=, $pop91, $pop344
+; CHECK-NEXT:    i32.or $push94=, $pop89, $pop93
+; CHECK-NEXT:    i8x16.extract_lane_u $push95=, $0, 4
+; CHECK-NEXT:    i32.const $push343=, 1
+; CHECK-NEXT:    i32.and $push96=, $pop95, $pop343
+; CHECK-NEXT:    i32.const $push342=, 4
+; CHECK-NEXT:    i32.shl $push98=, $pop96, $pop342
+; CHECK-NEXT:    i32.or $push99=, $pop94, $pop98
+; CHECK-NEXT:    i8x16.extract_lane_u $push100=, $0, 5
+; CHECK-NEXT:    i32.const $push341=, 1
+; CHECK-NEXT:    i32.and $push101=, $pop100, $pop341
+; CHECK-NEXT:    i32.const $push340=, 5
+; CHECK-NEXT:    i32.shl $push103=, $pop101, $pop340
+; CHECK-NEXT:    i32.or $push104=, $pop99, $pop103
+; CHECK-NEXT:    i8x16.extract_lane_u $push105=, $0, 6
+; CHECK-NEXT:    i32.const $push339=, 1
+; CHECK-NEXT:    i32.and $push106=, $pop105, $pop339
+; CHECK-NEXT:    i32.const $push338=, 6
+; CHECK-NEXT:    i32.shl $push108=, $pop106, $pop338
+; CHECK-NEXT:    i32.or $push109=, $pop104, $pop108
+; CHECK-NEXT:    i8x16.extract_lane_u $push110=, $0, 7
+; CHECK-NEXT:    i32.const $push337=, 1
+; CHECK-NEXT:    i32.and $push111=, $pop110, $pop337
+; CHECK-NEXT:    i32.const $push336=, 7
+; CHECK-NEXT:    i32.shl $push113=, $pop111, $pop336
+; CHECK-NEXT:    i32.or $push114=, $pop109, $pop113
+; CHECK-NEXT:    i8x16.extract_lane_u $push115=, $0, 8
+; CHECK-NEXT:    i32.const $push335=, 1
+; CHECK-NEXT:    i32.and $push116=, $pop115, $pop335
+; CHECK-NEXT:    i32.const $push334=, 8
+; CHECK-NEXT:    i32.shl $push118=, $pop116, $pop334
+; CHECK-NEXT:    i32.or $push119=, $pop114, $pop118
+; CHECK-NEXT:    i8x16.extract_lane_u $push120=, $0, 9
+; CHECK-NEXT:    i32.const $push333=, 1
+; CHECK-NEXT:    i32.and $push121=, $pop120, $pop333
+; CHECK-NEXT:    i32.const $push332=, 9
+; CHECK-NEXT:    i32.shl $push123=, $pop121, $pop332
+; CHECK-NEXT:    i32.or $push124=, $pop119, $pop123
+; CHECK-NEXT:    i8x16.extract_lane_u $push125=, $0, 10
+; CHECK-NEXT:    i32.const $push331=, 1
+; CHECK-NEXT:    i32.and $push126=, $pop125, $pop331
+; CHECK-NEXT:    i32.const $push330=, 10
+; CHECK-NEXT:    i32.shl $push128=, $pop126, $pop330
+; CHECK-NEXT:    i32.or $push129=, $pop124, $pop128
+; CHECK-NEXT:    i8x16.extract_lane_u $push130=, $0, 11
+; CHECK-NEXT:    i32.const $push329=, 1
+; CHECK-NEXT:    i32.and $push131=, $pop130, $pop329
+; CHECK-NEXT:    i32.const $push328=, 11
+; CHECK-NEXT:    i32.shl $push133=, $pop131, $pop328
+; CHECK-NEXT:    i32.or $push134=, $pop129, $pop133
+; CHECK-NEXT:    i8x16.extract_lane_u $push135=, $0, 12
+; CHECK-NEXT:    i32.const $push327=, 1
+; CHECK-NEXT:    i32.and $push136=, $pop135, $pop327
+; CHECK-NEXT:    i32.const $push326=, 12
+; CHECK-NEXT:    i32.shl $push138=, $pop136, $pop326
+; CHECK-NEXT:    i32.or $push139=, $pop134, $pop138
+; CHECK-NEXT:    i8x16.extract_lane_u $push140=, $0, 13
+; CHECK-NEXT:    i32.const $push325=, 1
+; CHECK-NEXT:    i32.and $push141=, $pop140, $pop325
+; CHECK-NEXT:    i32.const $push324=, 13
+; CHECK-NEXT:    i32.shl $push143=, $pop141, $pop324
+; CHECK-NEXT:    i32.or $push144=, $pop139, $pop143
+; CHECK-NEXT:    i8x16.extract_lane_u $push145=, $0, 14
+; CHECK-NEXT:    i32.const $push323=, 1
+; CHECK-NEXT:    i32.and $push146=, $pop145, $pop323
+; CHECK-NEXT:    i32.const $push322=, 14
+; CHECK-NEXT:    i32.shl $push148=, $pop146, $pop322
+; CHECK-NEXT:    i32.or $push149=, $pop144, $pop148
+; CHECK-NEXT:    i8x16.extract_lane_u $push150=, $0, 15
+; CHECK-NEXT:    i32.const $push321=, 15
+; CHECK-NEXT:    i32.shl $push152=, $pop150, $pop321
+; CHECK-NEXT:    i32.or $push153=, $pop149, $pop152
+; CHECK-NEXT:    i32.const $push320=, 65535
+; CHECK-NEXT:    i32.and $push155=, $pop153, $pop320
+; CHECK-NEXT:    i8x16.extract_lane_u $push75=, $1, 15
+; CHECK-NEXT:    i32.const $push319=, 31
+; CHECK-NEXT:    i32.shl $push77=, $pop75, $pop319
+; CHECK-NEXT:    i8x16.extract_lane_u $push70=, $1, 14
+; CHECK-NEXT:    i32.const $push318=, 1
+; CHECK-NEXT:    i32.and $push71=, $pop70, $pop318
+; CHECK-NEXT:    i32.const $push317=, 30
+; CHECK-NEXT:    i32.shl $push73=, $pop71, $pop317
+; CHECK-NEXT:    i8x16.extract_lane_u $push65=, $1, 13
+; CHECK-NEXT:    i32.const $push316=, 1
+; CHECK-NEXT:    i32.and $push66=, $pop65, $pop316
+; CHECK-NEXT:    i32.const $push315=, 29
+; CHECK-NEXT:    i32.shl $push68=, $pop66, $pop315
+; CHECK-NEXT:    i8x16.extract_lane_u $push60=, $1, 12
+; CHECK-NEXT:    i32.const $push314=, 1
+; CHECK-NEXT:    i32.and $push61=, $pop60, $pop314
+; CHECK-NEXT:    i32.const $push313=, 28
+; CHECK-NEXT:    i32.shl $push63=, $pop61, $pop313
+; CHECK-NEXT:    i8x16.extract_lane_u $push55=, $1, 11
+; CHECK-NEXT:    i32.const $push312=, 1
+; CHECK-NEXT:    i32.and $push56=, $pop55, $pop312
+; CHECK-NEXT:    i32.const $push311=, 27
+; CHECK-NEXT:    i32.shl $push58=, $pop56, $pop311
+; CHECK-NEXT:    i8x16.extract_lane_u $push50=, $1, 10
+; CHECK-NEXT:    i32.const $push310=, 1
+; CHECK-NEXT:    i32.and $push51=, $pop50, $pop310
+; CHECK-NEXT:    i32.const $push309=, 26
+; CHECK-NEXT:    i32.shl $push53=, $pop51, $pop309
+; CHECK-NEXT:    i8x16.extract_lane_u $push45=, $1, 9
+; CHECK-NEXT:    i32.const $push308=, 1
+; CHECK-NEXT:    i32.and $push46=, $pop45, $pop308
+; CHECK-NEXT:    i32.const $push307=, 25
+; CHECK-NEXT:    i32.shl $push48=, $pop46, $pop307
+; CHECK-NEXT:    i8x16.extract_lane_u $push40=, $1, 8
+; CHECK-NEXT:    i32.const $push306=, 1
+; CHECK-NEXT:    i32.and $push41=, $pop40, $pop306
+; CHECK-NEXT:    i32.const $push305=, 24
+; CHECK-NEXT:    i32.shl $push43=, $pop41, $pop305
+; CHECK-NEXT:    i8x16.extract_lane_u $push35=, $1, 7
+; CHECK-NEXT:    i32.const $push304=, 1
+; CHECK-NEXT:    i32.and $push36=, $pop35, $pop304
+; CHECK-NEXT:    i32.const $push303=, 23
+; CHECK-NEXT:    i32.shl $push38=, $pop36, $pop303
+; CHECK-NEXT:    i8x16.extract_lane_u $push30=, $1, 6
+; CHECK-NEXT:    i32.const $push302=, 1
+; CHECK-NEXT:    i32.and $push31=, $pop30, $pop302
+; CHECK-NEXT:    i32.const $push301=, 22
+; CHECK-NEXT:    i32.shl $push33=, $pop31, $pop301
+; CHECK-NEXT:    i8x16.extract_lane_u $push25=, $1, 5
+; CHECK-NEXT:    i32.const $push300=, 1
+; CHECK-NEXT:    i32.and $push26=, $pop25, $pop300
+; CHECK-NEXT:    i32.const $push299=, 21
+; CHECK-NEXT:    i32.shl $push28=, $pop26, $pop299
+; CHECK-NEXT:    i8x16.extract_lane_u $push20=, $1, 4
+; CHECK-NEXT:    i32.const $push298=, 1
+; CHECK-NEXT:    i32.and $push21=, $pop20, $pop298
+; CHECK-NEXT:    i32.const $push297=, 20
+; CHECK-NEXT:    i32.shl $push23=, $pop21, $pop297
+; CHECK-NEXT:    i8x16.extract_lane_u $push15=, $1, 3
+; CHECK-NEXT:    i32.const $push296=, 1
+; CHECK-NEXT:    i32.and $push16=, $pop15, $pop296
+; CHECK-NEXT:    i32.const $push295=, 19
+; CHECK-NEXT:    i32.shl $push18=, $pop16, $pop295
+; CHECK-NEXT:    i8x16.extract_lane_u $push10=, $1, 2
+; CHECK-NEXT:    i32.const $push294=, 1
+; CHECK-NEXT:    i32.and $push11=, $pop10, $pop294
+; CHECK-NEXT:    i32.const $push293=, 18
+; CHECK-NEXT:    i32.shl $push13=, $pop11, $pop293
+; CHECK-NEXT:    i8x16.extract_lane_u $push5=, $1, 1
+; CHECK-NEXT:    i32.const $push292=, 1
+; CHECK-NEXT:    i32.and $push6=, $pop5, $pop292
+; CHECK-NEXT:    i32.const $push291=, 17
+; CHECK-NEXT:    i32.shl $push8=, $pop6, $pop291
+; CHECK-NEXT:    i8x16.extract_lane_u $push0=, $1, 0
+; CHECK-NEXT:    i32.const $push290=, 1
+; CHECK-NEXT:    i32.and $push2=, $pop0, $pop290
+; CHECK-NEXT:    i32.const $push289=, 16
+; CHECK-NEXT:    i32.shl $push4=, $pop2, $pop289
+; CHECK-NEXT:    i32.or $push9=, $pop8, $pop4
+; CHECK-NEXT:    i32.or $push14=, $pop13, $pop9
+; CHECK-NEXT:    i32.or $push19=, $pop18, $pop14
+; CHECK-NEXT:    i32.or $push24=, $pop23, $pop19
+; CHECK-NEXT:    i32.or $push29=, $pop28, $pop24
+; CHECK-NEXT:    i32.or $push34=, $pop33, $pop29
+; CHECK-NEXT:    i32.or $push39=, $pop38, $pop34
+; CHECK-NEXT:    i32.or $push44=, $pop43, $pop39
+; CHECK-NEXT:    i32.or $push49=, $pop48, $pop44
+; CHECK-NEXT:    i32.or $push54=, $pop53, $pop49
+; CHECK-NEXT:    i32.or $push59=, $pop58, $pop54
+; CHECK-NEXT:    i32.or $push64=, $pop63, $pop59
+; CHECK-NEXT:    i32.or $push69=, $pop68, $pop64
+; CHECK-NEXT:    i32.or $push74=, $pop73, $pop69
+; CHECK-NEXT:    i32.or $push78=, $pop77, $pop74
+; CHECK-NEXT:    i32.or $push156=, $pop155, $pop78
+; CHECK-NEXT:    i64.extend_i32_u $push157=, $pop156
+; CHECK-NEXT:    i64.and $push284=, $pop283, $pop157
+; CHECK-NEXT:    i64.const $push285=, 4294967295
+; CHECK-NEXT:    i64.eq $push286=, $pop284, $pop285
+; CHECK-NEXT:    return $pop286
+  %bits = trunc <64 x i8> %x to <64 x i1>
+  %ret = call i1 @llvm.vector.reduce.and.v64i1(<64 x i1> %bits)
+  ret i1 %ret
+}
+
 ;; =====================
 ;; Test reduce after cmp
 ;; =====================
 
 define i1 @test_cmp_v16i8(<16 x i8> %x) {
 ; CHECK-LABEL: test_cmp_v16i8:
 ; CHECK:         .functype test_cmp_v16i8 (v128) -> (i32)
 ; CHECK-NEXT:  # %bb.0:
 ; CHECK-NEXT:    v128.const $push0=, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0
 ; CHECK-NEXT:    i8x16.eq $push1=, $0, $pop0
 ; CHECK-NEXT:    v128.any_true $push2=, $pop1
 ; CHECK-NEXT:    return $pop2
   %zero = icmp eq <16 x i8> %x, zeroinitializer
   %ret = call i1 @llvm.vector.reduce.or.v16i1(<16 x i1> %zero)
   ret i1 %ret
 }