[llvm] [AMDGPU] NFC, move testcase, only test output of promote-alloca with vector-combine (PR #166289)
via llvm-commits
llvm-commits at lists.llvm.org
Mon Nov 3 18:46:58 PST 2025
https://github.com/choikwa created https://github.com/llvm/llvm-project/pull/166289
None
>From 7014dca156df882a857b5f2fc89925b133696577 Mon Sep 17 00:00:00 2001
From: Kevin Choi <kevin.choi at amd.com>
Date: Thu, 30 Oct 2025 23:22:48 -0500
Subject: [PATCH 1/2] [AMDGPU] NFC, add testcase showing promote-alloca of
array of vectors to a large vector
---
.../AMDGPU/promote-alloca-array-to-vector.ll | 325 ++++++++++++++++++
1 file changed, 325 insertions(+)
create mode 100644 llvm/test/CodeGen/AMDGPU/promote-alloca-array-to-vector.ll
diff --git a/llvm/test/CodeGen/AMDGPU/promote-alloca-array-to-vector.ll b/llvm/test/CodeGen/AMDGPU/promote-alloca-array-to-vector.ll
new file mode 100644
index 0000000000000..05a0e39d4a715
--- /dev/null
+++ b/llvm/test/CodeGen/AMDGPU/promote-alloca-array-to-vector.ll
@@ -0,0 +1,325 @@
+; NOTE: Assertions have been autogenerated by utils/update_test_checks.py UTC_ARGS: --version 6
+; RUN: opt -S -mtriple=amdgcn-- -mcpu=gfx1100 -passes=amdgpu-promote-alloca < %s | FileCheck -check-prefix=OPT %s
+
+define amdgpu_kernel void @large_array_vectors_small_users(<16 x i8> %in, <16 x i8> %add, ptr addrspace(3) %out) #0 {
+; OPT-LABEL: define amdgpu_kernel void @large_array_vectors_small_users(
+; OPT-SAME: <16 x i8> [[IN:%.*]], <16 x i8> [[ADD:%.*]], ptr addrspace(3) [[OUT:%.*]]) #[[ATTR0:[0-9]+]] {
+; OPT-NEXT: [[ENTRY:.*:]]
+; OPT-NEXT: [[ALLOCA:%.*]] = freeze <128 x i8> poison
+; OPT-NEXT: [[TMP0:%.*]] = extractelement <16 x i8> [[IN]], i64 0
+; OPT-NEXT: [[TMP1:%.*]] = insertelement <128 x i8> [[ALLOCA]], i8 [[TMP0]], i32 0
+; OPT-NEXT: [[TMP2:%.*]] = extractelement <16 x i8> [[IN]], i64 1
+; OPT-NEXT: [[TMP3:%.*]] = insertelement <128 x i8> [[TMP1]], i8 [[TMP2]], i32 1
+; OPT-NEXT: [[TMP4:%.*]] = extractelement <16 x i8> [[IN]], i64 2
+; OPT-NEXT: [[TMP5:%.*]] = insertelement <128 x i8> [[TMP3]], i8 [[TMP4]], i32 2
+; OPT-NEXT: [[TMP6:%.*]] = extractelement <16 x i8> [[IN]], i64 3
+; OPT-NEXT: [[TMP7:%.*]] = insertelement <128 x i8> [[TMP5]], i8 [[TMP6]], i32 3
+; OPT-NEXT: [[TMP8:%.*]] = extractelement <16 x i8> [[IN]], i64 4
+; OPT-NEXT: [[TMP9:%.*]] = insertelement <128 x i8> [[TMP7]], i8 [[TMP8]], i32 4
+; OPT-NEXT: [[TMP10:%.*]] = extractelement <16 x i8> [[IN]], i64 5
+; OPT-NEXT: [[TMP11:%.*]] = insertelement <128 x i8> [[TMP9]], i8 [[TMP10]], i32 5
+; OPT-NEXT: [[TMP12:%.*]] = extractelement <16 x i8> [[IN]], i64 6
+; OPT-NEXT: [[TMP13:%.*]] = insertelement <128 x i8> [[TMP11]], i8 [[TMP12]], i32 6
+; OPT-NEXT: [[TMP14:%.*]] = extractelement <16 x i8> [[IN]], i64 7
+; OPT-NEXT: [[TMP15:%.*]] = insertelement <128 x i8> [[TMP13]], i8 [[TMP14]], i32 7
+; OPT-NEXT: [[TMP16:%.*]] = extractelement <16 x i8> [[IN]], i64 8
+; OPT-NEXT: [[TMP17:%.*]] = insertelement <128 x i8> [[TMP15]], i8 [[TMP16]], i32 8
+; OPT-NEXT: [[TMP18:%.*]] = extractelement <16 x i8> [[IN]], i64 9
+; OPT-NEXT: [[TMP19:%.*]] = insertelement <128 x i8> [[TMP17]], i8 [[TMP18]], i32 9
+; OPT-NEXT: [[TMP20:%.*]] = extractelement <16 x i8> [[IN]], i64 10
+; OPT-NEXT: [[TMP21:%.*]] = insertelement <128 x i8> [[TMP19]], i8 [[TMP20]], i32 10
+; OPT-NEXT: [[TMP22:%.*]] = extractelement <16 x i8> [[IN]], i64 11
+; OPT-NEXT: [[TMP23:%.*]] = insertelement <128 x i8> [[TMP21]], i8 [[TMP22]], i32 11
+; OPT-NEXT: [[TMP24:%.*]] = extractelement <16 x i8> [[IN]], i64 12
+; OPT-NEXT: [[TMP25:%.*]] = insertelement <128 x i8> [[TMP23]], i8 [[TMP24]], i32 12
+; OPT-NEXT: [[TMP26:%.*]] = extractelement <16 x i8> [[IN]], i64 13
+; OPT-NEXT: [[TMP27:%.*]] = insertelement <128 x i8> [[TMP25]], i8 [[TMP26]], i32 13
+; OPT-NEXT: [[TMP28:%.*]] = extractelement <16 x i8> [[IN]], i64 14
+; OPT-NEXT: [[TMP29:%.*]] = insertelement <128 x i8> [[TMP27]], i8 [[TMP28]], i32 14
+; OPT-NEXT: [[TMP30:%.*]] = extractelement <16 x i8> [[IN]], i64 15
+; OPT-NEXT: [[TMP31:%.*]] = insertelement <128 x i8> [[TMP29]], i8 [[TMP30]], i32 15
+; OPT-NEXT: [[TMP32:%.*]] = extractelement <16 x i8> [[IN]], i64 0
+; OPT-NEXT: [[TMP33:%.*]] = insertelement <128 x i8> [[TMP31]], i8 [[TMP32]], i32 0
+; OPT-NEXT: [[TMP34:%.*]] = extractelement <16 x i8> [[IN]], i64 1
+; OPT-NEXT: [[TMP35:%.*]] = insertelement <128 x i8> [[TMP33]], i8 [[TMP34]], i32 1
+; OPT-NEXT: [[TMP36:%.*]] = extractelement <16 x i8> [[IN]], i64 2
+; OPT-NEXT: [[TMP37:%.*]] = insertelement <128 x i8> [[TMP35]], i8 [[TMP36]], i32 2
+; OPT-NEXT: [[TMP38:%.*]] = extractelement <16 x i8> [[IN]], i64 3
+; OPT-NEXT: [[TMP39:%.*]] = insertelement <128 x i8> [[TMP37]], i8 [[TMP38]], i32 3
+; OPT-NEXT: [[TMP40:%.*]] = extractelement <16 x i8> [[IN]], i64 4
+; OPT-NEXT: [[TMP41:%.*]] = insertelement <128 x i8> [[TMP39]], i8 [[TMP40]], i32 4
+; OPT-NEXT: [[TMP42:%.*]] = extractelement <16 x i8> [[IN]], i64 5
+; OPT-NEXT: [[TMP43:%.*]] = insertelement <128 x i8> [[TMP41]], i8 [[TMP42]], i32 5
+; OPT-NEXT: [[TMP44:%.*]] = extractelement <16 x i8> [[IN]], i64 6
+; OPT-NEXT: [[TMP45:%.*]] = insertelement <128 x i8> [[TMP43]], i8 [[TMP44]], i32 6
+; OPT-NEXT: [[TMP46:%.*]] = extractelement <16 x i8> [[IN]], i64 7
+; OPT-NEXT: [[TMP47:%.*]] = insertelement <128 x i8> [[TMP45]], i8 [[TMP46]], i32 7
+; OPT-NEXT: [[TMP48:%.*]] = extractelement <16 x i8> [[IN]], i64 8
+; OPT-NEXT: [[TMP49:%.*]] = insertelement <128 x i8> [[TMP47]], i8 [[TMP48]], i32 8
+; OPT-NEXT: [[TMP50:%.*]] = extractelement <16 x i8> [[IN]], i64 9
+; OPT-NEXT: [[TMP51:%.*]] = insertelement <128 x i8> [[TMP49]], i8 [[TMP50]], i32 9
+; OPT-NEXT: [[TMP52:%.*]] = extractelement <16 x i8> [[IN]], i64 10
+; OPT-NEXT: [[TMP53:%.*]] = insertelement <128 x i8> [[TMP51]], i8 [[TMP52]], i32 10
+; OPT-NEXT: [[TMP54:%.*]] = extractelement <16 x i8> [[IN]], i64 11
+; OPT-NEXT: [[TMP55:%.*]] = insertelement <128 x i8> [[TMP53]], i8 [[TMP54]], i32 11
+; OPT-NEXT: [[TMP56:%.*]] = extractelement <16 x i8> [[IN]], i64 12
+; OPT-NEXT: [[TMP57:%.*]] = insertelement <128 x i8> [[TMP55]], i8 [[TMP56]], i32 12
+; OPT-NEXT: [[TMP58:%.*]] = extractelement <16 x i8> [[IN]], i64 13
+; OPT-NEXT: [[TMP59:%.*]] = insertelement <128 x i8> [[TMP57]], i8 [[TMP58]], i32 13
+; OPT-NEXT: [[TMP60:%.*]] = extractelement <16 x i8> [[IN]], i64 14
+; OPT-NEXT: [[TMP61:%.*]] = insertelement <128 x i8> [[TMP59]], i8 [[TMP60]], i32 14
+; OPT-NEXT: [[TMP62:%.*]] = extractelement <16 x i8> [[IN]], i64 15
+; OPT-NEXT: [[TMP63:%.*]] = insertelement <128 x i8> [[TMP61]], i8 [[TMP62]], i32 15
+; OPT-NEXT: [[TMP64:%.*]] = extractelement <16 x i8> [[IN]], i64 0
+; OPT-NEXT: [[TMP65:%.*]] = insertelement <128 x i8> [[TMP63]], i8 [[TMP64]], i32 0
+; OPT-NEXT: [[TMP66:%.*]] = extractelement <16 x i8> [[IN]], i64 1
+; OPT-NEXT: [[TMP67:%.*]] = insertelement <128 x i8> [[TMP65]], i8 [[TMP66]], i32 1
+; OPT-NEXT: [[TMP68:%.*]] = extractelement <16 x i8> [[IN]], i64 2
+; OPT-NEXT: [[TMP69:%.*]] = insertelement <128 x i8> [[TMP67]], i8 [[TMP68]], i32 2
+; OPT-NEXT: [[TMP70:%.*]] = extractelement <16 x i8> [[IN]], i64 3
+; OPT-NEXT: [[TMP71:%.*]] = insertelement <128 x i8> [[TMP69]], i8 [[TMP70]], i32 3
+; OPT-NEXT: [[TMP72:%.*]] = extractelement <16 x i8> [[IN]], i64 4
+; OPT-NEXT: [[TMP73:%.*]] = insertelement <128 x i8> [[TMP71]], i8 [[TMP72]], i32 4
+; OPT-NEXT: [[TMP74:%.*]] = extractelement <16 x i8> [[IN]], i64 5
+; OPT-NEXT: [[TMP75:%.*]] = insertelement <128 x i8> [[TMP73]], i8 [[TMP74]], i32 5
+; OPT-NEXT: [[TMP76:%.*]] = extractelement <16 x i8> [[IN]], i64 6
+; OPT-NEXT: [[TMP77:%.*]] = insertelement <128 x i8> [[TMP75]], i8 [[TMP76]], i32 6
+; OPT-NEXT: [[TMP78:%.*]] = extractelement <16 x i8> [[IN]], i64 7
+; OPT-NEXT: [[TMP79:%.*]] = insertelement <128 x i8> [[TMP77]], i8 [[TMP78]], i32 7
+; OPT-NEXT: [[TMP80:%.*]] = extractelement <16 x i8> [[IN]], i64 8
+; OPT-NEXT: [[TMP81:%.*]] = insertelement <128 x i8> [[TMP79]], i8 [[TMP80]], i32 8
+; OPT-NEXT: [[TMP82:%.*]] = extractelement <16 x i8> [[IN]], i64 9
+; OPT-NEXT: [[TMP83:%.*]] = insertelement <128 x i8> [[TMP81]], i8 [[TMP82]], i32 9
+; OPT-NEXT: [[TMP84:%.*]] = extractelement <16 x i8> [[IN]], i64 10
+; OPT-NEXT: [[TMP85:%.*]] = insertelement <128 x i8> [[TMP83]], i8 [[TMP84]], i32 10
+; OPT-NEXT: [[TMP86:%.*]] = extractelement <16 x i8> [[IN]], i64 11
+; OPT-NEXT: [[TMP87:%.*]] = insertelement <128 x i8> [[TMP85]], i8 [[TMP86]], i32 11
+; OPT-NEXT: [[TMP88:%.*]] = extractelement <16 x i8> [[IN]], i64 12
+; OPT-NEXT: [[TMP89:%.*]] = insertelement <128 x i8> [[TMP87]], i8 [[TMP88]], i32 12
+; OPT-NEXT: [[TMP90:%.*]] = extractelement <16 x i8> [[IN]], i64 13
+; OPT-NEXT: [[TMP91:%.*]] = insertelement <128 x i8> [[TMP89]], i8 [[TMP90]], i32 13
+; OPT-NEXT: [[TMP92:%.*]] = extractelement <16 x i8> [[IN]], i64 14
+; OPT-NEXT: [[TMP93:%.*]] = insertelement <128 x i8> [[TMP91]], i8 [[TMP92]], i32 14
+; OPT-NEXT: [[TMP94:%.*]] = extractelement <16 x i8> [[IN]], i64 15
+; OPT-NEXT: [[TMP95:%.*]] = insertelement <128 x i8> [[TMP93]], i8 [[TMP94]], i32 15
+; OPT-NEXT: [[TMP96:%.*]] = extractelement <16 x i8> [[IN]], i64 0
+; OPT-NEXT: [[TMP97:%.*]] = insertelement <128 x i8> [[TMP95]], i8 [[TMP96]], i32 0
+; OPT-NEXT: [[TMP98:%.*]] = extractelement <16 x i8> [[IN]], i64 1
+; OPT-NEXT: [[TMP99:%.*]] = insertelement <128 x i8> [[TMP97]], i8 [[TMP98]], i32 1
+; OPT-NEXT: [[TMP100:%.*]] = extractelement <16 x i8> [[IN]], i64 2
+; OPT-NEXT: [[TMP101:%.*]] = insertelement <128 x i8> [[TMP99]], i8 [[TMP100]], i32 2
+; OPT-NEXT: [[TMP102:%.*]] = extractelement <16 x i8> [[IN]], i64 3
+; OPT-NEXT: [[TMP103:%.*]] = insertelement <128 x i8> [[TMP101]], i8 [[TMP102]], i32 3
+; OPT-NEXT: [[TMP104:%.*]] = extractelement <16 x i8> [[IN]], i64 4
+; OPT-NEXT: [[TMP105:%.*]] = insertelement <128 x i8> [[TMP103]], i8 [[TMP104]], i32 4
+; OPT-NEXT: [[TMP106:%.*]] = extractelement <16 x i8> [[IN]], i64 5
+; OPT-NEXT: [[TMP107:%.*]] = insertelement <128 x i8> [[TMP105]], i8 [[TMP106]], i32 5
+; OPT-NEXT: [[TMP108:%.*]] = extractelement <16 x i8> [[IN]], i64 6
+; OPT-NEXT: [[TMP109:%.*]] = insertelement <128 x i8> [[TMP107]], i8 [[TMP108]], i32 6
+; OPT-NEXT: [[TMP110:%.*]] = extractelement <16 x i8> [[IN]], i64 7
+; OPT-NEXT: [[TMP111:%.*]] = insertelement <128 x i8> [[TMP109]], i8 [[TMP110]], i32 7
+; OPT-NEXT: [[TMP112:%.*]] = extractelement <16 x i8> [[IN]], i64 8
+; OPT-NEXT: [[TMP113:%.*]] = insertelement <128 x i8> [[TMP111]], i8 [[TMP112]], i32 8
+; OPT-NEXT: [[TMP114:%.*]] = extractelement <16 x i8> [[IN]], i64 9
+; OPT-NEXT: [[TMP115:%.*]] = insertelement <128 x i8> [[TMP113]], i8 [[TMP114]], i32 9
+; OPT-NEXT: [[TMP116:%.*]] = extractelement <16 x i8> [[IN]], i64 10
+; OPT-NEXT: [[TMP117:%.*]] = insertelement <128 x i8> [[TMP115]], i8 [[TMP116]], i32 10
+; OPT-NEXT: [[TMP118:%.*]] = extractelement <16 x i8> [[IN]], i64 11
+; OPT-NEXT: [[TMP119:%.*]] = insertelement <128 x i8> [[TMP117]], i8 [[TMP118]], i32 11
+; OPT-NEXT: [[TMP120:%.*]] = extractelement <16 x i8> [[IN]], i64 12
+; OPT-NEXT: [[TMP121:%.*]] = insertelement <128 x i8> [[TMP119]], i8 [[TMP120]], i32 12
+; OPT-NEXT: [[TMP122:%.*]] = extractelement <16 x i8> [[IN]], i64 13
+; OPT-NEXT: [[TMP123:%.*]] = insertelement <128 x i8> [[TMP121]], i8 [[TMP122]], i32 13
+; OPT-NEXT: [[TMP124:%.*]] = extractelement <16 x i8> [[IN]], i64 14
+; OPT-NEXT: [[TMP125:%.*]] = insertelement <128 x i8> [[TMP123]], i8 [[TMP124]], i32 14
+; OPT-NEXT: [[TMP126:%.*]] = extractelement <16 x i8> [[IN]], i64 15
+; OPT-NEXT: [[TMP127:%.*]] = insertelement <128 x i8> [[TMP125]], i8 [[TMP126]], i32 15
+; OPT-NEXT: [[TMP128:%.*]] = extractelement <16 x i8> [[IN]], i64 0
+; OPT-NEXT: [[TMP129:%.*]] = insertelement <128 x i8> [[TMP127]], i8 [[TMP128]], i32 0
+; OPT-NEXT: [[TMP130:%.*]] = extractelement <16 x i8> [[IN]], i64 1
+; OPT-NEXT: [[TMP131:%.*]] = insertelement <128 x i8> [[TMP129]], i8 [[TMP130]], i32 1
+; OPT-NEXT: [[TMP132:%.*]] = extractelement <16 x i8> [[IN]], i64 2
+; OPT-NEXT: [[TMP133:%.*]] = insertelement <128 x i8> [[TMP131]], i8 [[TMP132]], i32 2
+; OPT-NEXT: [[TMP134:%.*]] = extractelement <16 x i8> [[IN]], i64 3
+; OPT-NEXT: [[TMP135:%.*]] = insertelement <128 x i8> [[TMP133]], i8 [[TMP134]], i32 3
+; OPT-NEXT: [[TMP136:%.*]] = extractelement <16 x i8> [[IN]], i64 4
+; OPT-NEXT: [[TMP137:%.*]] = insertelement <128 x i8> [[TMP135]], i8 [[TMP136]], i32 4
+; OPT-NEXT: [[TMP138:%.*]] = extractelement <16 x i8> [[IN]], i64 5
+; OPT-NEXT: [[TMP139:%.*]] = insertelement <128 x i8> [[TMP137]], i8 [[TMP138]], i32 5
+; OPT-NEXT: [[TMP140:%.*]] = extractelement <16 x i8> [[IN]], i64 6
+; OPT-NEXT: [[TMP141:%.*]] = insertelement <128 x i8> [[TMP139]], i8 [[TMP140]], i32 6
+; OPT-NEXT: [[TMP142:%.*]] = extractelement <16 x i8> [[IN]], i64 7
+; OPT-NEXT: [[TMP143:%.*]] = insertelement <128 x i8> [[TMP141]], i8 [[TMP142]], i32 7
+; OPT-NEXT: [[TMP144:%.*]] = extractelement <16 x i8> [[IN]], i64 8
+; OPT-NEXT: [[TMP145:%.*]] = insertelement <128 x i8> [[TMP143]], i8 [[TMP144]], i32 8
+; OPT-NEXT: [[TMP146:%.*]] = extractelement <16 x i8> [[IN]], i64 9
+; OPT-NEXT: [[TMP147:%.*]] = insertelement <128 x i8> [[TMP145]], i8 [[TMP146]], i32 9
+; OPT-NEXT: [[TMP148:%.*]] = extractelement <16 x i8> [[IN]], i64 10
+; OPT-NEXT: [[TMP149:%.*]] = insertelement <128 x i8> [[TMP147]], i8 [[TMP148]], i32 10
+; OPT-NEXT: [[TMP150:%.*]] = extractelement <16 x i8> [[IN]], i64 11
+; OPT-NEXT: [[TMP151:%.*]] = insertelement <128 x i8> [[TMP149]], i8 [[TMP150]], i32 11
+; OPT-NEXT: [[TMP152:%.*]] = extractelement <16 x i8> [[IN]], i64 12
+; OPT-NEXT: [[TMP153:%.*]] = insertelement <128 x i8> [[TMP151]], i8 [[TMP152]], i32 12
+; OPT-NEXT: [[TMP154:%.*]] = extractelement <16 x i8> [[IN]], i64 13
+; OPT-NEXT: [[TMP155:%.*]] = insertelement <128 x i8> [[TMP153]], i8 [[TMP154]], i32 13
+; OPT-NEXT: [[TMP156:%.*]] = extractelement <16 x i8> [[IN]], i64 14
+; OPT-NEXT: [[TMP157:%.*]] = insertelement <128 x i8> [[TMP155]], i8 [[TMP156]], i32 14
+; OPT-NEXT: [[TMP158:%.*]] = extractelement <16 x i8> [[IN]], i64 15
+; OPT-NEXT: [[TMP159:%.*]] = insertelement <128 x i8> [[TMP157]], i8 [[TMP158]], i32 15
+; OPT-NEXT: [[TMP160:%.*]] = extractelement <16 x i8> [[IN]], i64 0
+; OPT-NEXT: [[TMP161:%.*]] = insertelement <128 x i8> [[TMP159]], i8 [[TMP160]], i32 0
+; OPT-NEXT: [[TMP162:%.*]] = extractelement <16 x i8> [[IN]], i64 1
+; OPT-NEXT: [[TMP163:%.*]] = insertelement <128 x i8> [[TMP161]], i8 [[TMP162]], i32 1
+; OPT-NEXT: [[TMP164:%.*]] = extractelement <16 x i8> [[IN]], i64 2
+; OPT-NEXT: [[TMP165:%.*]] = insertelement <128 x i8> [[TMP163]], i8 [[TMP164]], i32 2
+; OPT-NEXT: [[TMP166:%.*]] = extractelement <16 x i8> [[IN]], i64 3
+; OPT-NEXT: [[TMP167:%.*]] = insertelement <128 x i8> [[TMP165]], i8 [[TMP166]], i32 3
+; OPT-NEXT: [[TMP168:%.*]] = extractelement <16 x i8> [[IN]], i64 4
+; OPT-NEXT: [[TMP169:%.*]] = insertelement <128 x i8> [[TMP167]], i8 [[TMP168]], i32 4
+; OPT-NEXT: [[TMP170:%.*]] = extractelement <16 x i8> [[IN]], i64 5
+; OPT-NEXT: [[TMP171:%.*]] = insertelement <128 x i8> [[TMP169]], i8 [[TMP170]], i32 5
+; OPT-NEXT: [[TMP172:%.*]] = extractelement <16 x i8> [[IN]], i64 6
+; OPT-NEXT: [[TMP173:%.*]] = insertelement <128 x i8> [[TMP171]], i8 [[TMP172]], i32 6
+; OPT-NEXT: [[TMP174:%.*]] = extractelement <16 x i8> [[IN]], i64 7
+; OPT-NEXT: [[TMP175:%.*]] = insertelement <128 x i8> [[TMP173]], i8 [[TMP174]], i32 7
+; OPT-NEXT: [[TMP176:%.*]] = extractelement <16 x i8> [[IN]], i64 8
+; OPT-NEXT: [[TMP177:%.*]] = insertelement <128 x i8> [[TMP175]], i8 [[TMP176]], i32 8
+; OPT-NEXT: [[TMP178:%.*]] = extractelement <16 x i8> [[IN]], i64 9
+; OPT-NEXT: [[TMP179:%.*]] = insertelement <128 x i8> [[TMP177]], i8 [[TMP178]], i32 9
+; OPT-NEXT: [[TMP180:%.*]] = extractelement <16 x i8> [[IN]], i64 10
+; OPT-NEXT: [[TMP181:%.*]] = insertelement <128 x i8> [[TMP179]], i8 [[TMP180]], i32 10
+; OPT-NEXT: [[TMP182:%.*]] = extractelement <16 x i8> [[IN]], i64 11
+; OPT-NEXT: [[TMP183:%.*]] = insertelement <128 x i8> [[TMP181]], i8 [[TMP182]], i32 11
+; OPT-NEXT: [[TMP184:%.*]] = extractelement <16 x i8> [[IN]], i64 12
+; OPT-NEXT: [[TMP185:%.*]] = insertelement <128 x i8> [[TMP183]], i8 [[TMP184]], i32 12
+; OPT-NEXT: [[TMP186:%.*]] = extractelement <16 x i8> [[IN]], i64 13
+; OPT-NEXT: [[TMP187:%.*]] = insertelement <128 x i8> [[TMP185]], i8 [[TMP186]], i32 13
+; OPT-NEXT: [[TMP188:%.*]] = extractelement <16 x i8> [[IN]], i64 14
+; OPT-NEXT: [[TMP189:%.*]] = insertelement <128 x i8> [[TMP187]], i8 [[TMP188]], i32 14
+; OPT-NEXT: [[TMP190:%.*]] = extractelement <16 x i8> [[IN]], i64 15
+; OPT-NEXT: [[TMP191:%.*]] = insertelement <128 x i8> [[TMP189]], i8 [[TMP190]], i32 15
+; OPT-NEXT: [[TMP192:%.*]] = extractelement <16 x i8> [[IN]], i64 0
+; OPT-NEXT: [[TMP193:%.*]] = insertelement <128 x i8> [[TMP191]], i8 [[TMP192]], i32 0
+; OPT-NEXT: [[TMP194:%.*]] = extractelement <16 x i8> [[IN]], i64 1
+; OPT-NEXT: [[TMP195:%.*]] = insertelement <128 x i8> [[TMP193]], i8 [[TMP194]], i32 1
+; OPT-NEXT: [[TMP196:%.*]] = extractelement <16 x i8> [[IN]], i64 2
+; OPT-NEXT: [[TMP197:%.*]] = insertelement <128 x i8> [[TMP195]], i8 [[TMP196]], i32 2
+; OPT-NEXT: [[TMP198:%.*]] = extractelement <16 x i8> [[IN]], i64 3
+; OPT-NEXT: [[TMP199:%.*]] = insertelement <128 x i8> [[TMP197]], i8 [[TMP198]], i32 3
+; OPT-NEXT: [[TMP200:%.*]] = extractelement <16 x i8> [[IN]], i64 4
+; OPT-NEXT: [[TMP201:%.*]] = insertelement <128 x i8> [[TMP199]], i8 [[TMP200]], i32 4
+; OPT-NEXT: [[TMP202:%.*]] = extractelement <16 x i8> [[IN]], i64 5
+; OPT-NEXT: [[TMP203:%.*]] = insertelement <128 x i8> [[TMP201]], i8 [[TMP202]], i32 5
+; OPT-NEXT: [[TMP204:%.*]] = extractelement <16 x i8> [[IN]], i64 6
+; OPT-NEXT: [[TMP205:%.*]] = insertelement <128 x i8> [[TMP203]], i8 [[TMP204]], i32 6
+; OPT-NEXT: [[TMP206:%.*]] = extractelement <16 x i8> [[IN]], i64 7
+; OPT-NEXT: [[TMP207:%.*]] = insertelement <128 x i8> [[TMP205]], i8 [[TMP206]], i32 7
+; OPT-NEXT: [[TMP208:%.*]] = extractelement <16 x i8> [[IN]], i64 8
+; OPT-NEXT: [[TMP209:%.*]] = insertelement <128 x i8> [[TMP207]], i8 [[TMP208]], i32 8
+; OPT-NEXT: [[TMP210:%.*]] = extractelement <16 x i8> [[IN]], i64 9
+; OPT-NEXT: [[TMP211:%.*]] = insertelement <128 x i8> [[TMP209]], i8 [[TMP210]], i32 9
+; OPT-NEXT: [[TMP212:%.*]] = extractelement <16 x i8> [[IN]], i64 10
+; OPT-NEXT: [[TMP213:%.*]] = insertelement <128 x i8> [[TMP211]], i8 [[TMP212]], i32 10
+; OPT-NEXT: [[TMP214:%.*]] = extractelement <16 x i8> [[IN]], i64 11
+; OPT-NEXT: [[TMP215:%.*]] = insertelement <128 x i8> [[TMP213]], i8 [[TMP214]], i32 11
+; OPT-NEXT: [[TMP216:%.*]] = extractelement <16 x i8> [[IN]], i64 12
+; OPT-NEXT: [[TMP217:%.*]] = insertelement <128 x i8> [[TMP215]], i8 [[TMP216]], i32 12
+; OPT-NEXT: [[TMP218:%.*]] = extractelement <16 x i8> [[IN]], i64 13
+; OPT-NEXT: [[TMP219:%.*]] = insertelement <128 x i8> [[TMP217]], i8 [[TMP218]], i32 13
+; OPT-NEXT: [[TMP220:%.*]] = extractelement <16 x i8> [[IN]], i64 14
+; OPT-NEXT: [[TMP221:%.*]] = insertelement <128 x i8> [[TMP219]], i8 [[TMP220]], i32 14
+; OPT-NEXT: [[TMP222:%.*]] = extractelement <16 x i8> [[IN]], i64 15
+; OPT-NEXT: [[TMP223:%.*]] = insertelement <128 x i8> [[TMP221]], i8 [[TMP222]], i32 15
+; OPT-NEXT: [[TMP224:%.*]] = extractelement <16 x i8> [[IN]], i64 0
+; OPT-NEXT: [[TMP225:%.*]] = insertelement <128 x i8> [[TMP223]], i8 [[TMP224]], i32 0
+; OPT-NEXT: [[TMP226:%.*]] = extractelement <16 x i8> [[IN]], i64 1
+; OPT-NEXT: [[TMP227:%.*]] = insertelement <128 x i8> [[TMP225]], i8 [[TMP226]], i32 1
+; OPT-NEXT: [[TMP228:%.*]] = extractelement <16 x i8> [[IN]], i64 2
+; OPT-NEXT: [[TMP229:%.*]] = insertelement <128 x i8> [[TMP227]], i8 [[TMP228]], i32 2
+; OPT-NEXT: [[TMP230:%.*]] = extractelement <16 x i8> [[IN]], i64 3
+; OPT-NEXT: [[TMP231:%.*]] = insertelement <128 x i8> [[TMP229]], i8 [[TMP230]], i32 3
+; OPT-NEXT: [[TMP232:%.*]] = extractelement <16 x i8> [[IN]], i64 4
+; OPT-NEXT: [[TMP233:%.*]] = insertelement <128 x i8> [[TMP231]], i8 [[TMP232]], i32 4
+; OPT-NEXT: [[TMP234:%.*]] = extractelement <16 x i8> [[IN]], i64 5
+; OPT-NEXT: [[TMP235:%.*]] = insertelement <128 x i8> [[TMP233]], i8 [[TMP234]], i32 5
+; OPT-NEXT: [[TMP236:%.*]] = extractelement <16 x i8> [[IN]], i64 6
+; OPT-NEXT: [[TMP237:%.*]] = insertelement <128 x i8> [[TMP235]], i8 [[TMP236]], i32 6
+; OPT-NEXT: [[TMP238:%.*]] = extractelement <16 x i8> [[IN]], i64 7
+; OPT-NEXT: [[TMP239:%.*]] = insertelement <128 x i8> [[TMP237]], i8 [[TMP238]], i32 7
+; OPT-NEXT: [[TMP240:%.*]] = extractelement <16 x i8> [[IN]], i64 8
+; OPT-NEXT: [[TMP241:%.*]] = insertelement <128 x i8> [[TMP239]], i8 [[TMP240]], i32 8
+; OPT-NEXT: [[TMP242:%.*]] = extractelement <16 x i8> [[IN]], i64 9
+; OPT-NEXT: [[TMP243:%.*]] = insertelement <128 x i8> [[TMP241]], i8 [[TMP242]], i32 9
+; OPT-NEXT: [[TMP244:%.*]] = extractelement <16 x i8> [[IN]], i64 10
+; OPT-NEXT: [[TMP245:%.*]] = insertelement <128 x i8> [[TMP243]], i8 [[TMP244]], i32 10
+; OPT-NEXT: [[TMP246:%.*]] = extractelement <16 x i8> [[IN]], i64 11
+; OPT-NEXT: [[TMP247:%.*]] = insertelement <128 x i8> [[TMP245]], i8 [[TMP246]], i32 11
+; OPT-NEXT: [[TMP248:%.*]] = extractelement <16 x i8> [[IN]], i64 12
+; OPT-NEXT: [[TMP249:%.*]] = insertelement <128 x i8> [[TMP247]], i8 [[TMP248]], i32 12
+; OPT-NEXT: [[TMP250:%.*]] = extractelement <16 x i8> [[IN]], i64 13
+; OPT-NEXT: [[TMP251:%.*]] = insertelement <128 x i8> [[TMP249]], i8 [[TMP250]], i32 13
+; OPT-NEXT: [[TMP252:%.*]] = extractelement <16 x i8> [[IN]], i64 14
+; OPT-NEXT: [[TMP253:%.*]] = insertelement <128 x i8> [[TMP251]], i8 [[TMP252]], i32 14
+; OPT-NEXT: [[TMP254:%.*]] = extractelement <16 x i8> [[IN]], i64 15
+; OPT-NEXT: [[TMP255:%.*]] = insertelement <128 x i8> [[TMP253]], i8 [[TMP254]], i32 15
+; OPT-NEXT: [[TMP256:%.*]] = extractelement <128 x i8> [[TMP255]], i32 80
+; OPT-NEXT: [[TMP257:%.*]] = insertelement <16 x i8> poison, i8 [[TMP256]], i64 0
+; OPT-NEXT: [[TMP258:%.*]] = extractelement <128 x i8> [[TMP255]], i32 81
+; OPT-NEXT: [[TMP259:%.*]] = insertelement <16 x i8> [[TMP257]], i8 [[TMP258]], i64 1
+; OPT-NEXT: [[TMP260:%.*]] = extractelement <128 x i8> [[TMP255]], i32 82
+; OPT-NEXT: [[TMP261:%.*]] = insertelement <16 x i8> [[TMP259]], i8 [[TMP260]], i64 2
+; OPT-NEXT: [[TMP262:%.*]] = extractelement <128 x i8> [[TMP255]], i32 83
+; OPT-NEXT: [[TMP263:%.*]] = insertelement <16 x i8> [[TMP261]], i8 [[TMP262]], i64 3
+; OPT-NEXT: [[TMP264:%.*]] = extractelement <128 x i8> [[TMP255]], i32 84
+; OPT-NEXT: [[TMP265:%.*]] = insertelement <16 x i8> [[TMP263]], i8 [[TMP264]], i64 4
+; OPT-NEXT: [[TMP266:%.*]] = extractelement <128 x i8> [[TMP255]], i32 85
+; OPT-NEXT: [[TMP267:%.*]] = insertelement <16 x i8> [[TMP265]], i8 [[TMP266]], i64 5
+; OPT-NEXT: [[TMP268:%.*]] = extractelement <128 x i8> [[TMP255]], i32 86
+; OPT-NEXT: [[TMP269:%.*]] = insertelement <16 x i8> [[TMP267]], i8 [[TMP268]], i64 6
+; OPT-NEXT: [[TMP270:%.*]] = extractelement <128 x i8> [[TMP255]], i32 87
+; OPT-NEXT: [[TMP271:%.*]] = insertelement <16 x i8> [[TMP269]], i8 [[TMP270]], i64 7
+; OPT-NEXT: [[TMP272:%.*]] = extractelement <128 x i8> [[TMP255]], i32 88
+; OPT-NEXT: [[TMP273:%.*]] = insertelement <16 x i8> [[TMP271]], i8 [[TMP272]], i64 8
+; OPT-NEXT: [[TMP274:%.*]] = extractelement <128 x i8> [[TMP255]], i32 89
+; OPT-NEXT: [[TMP275:%.*]] = insertelement <16 x i8> [[TMP273]], i8 [[TMP274]], i64 9
+; OPT-NEXT: [[TMP276:%.*]] = extractelement <128 x i8> [[TMP255]], i32 90
+; OPT-NEXT: [[TMP277:%.*]] = insertelement <16 x i8> [[TMP275]], i8 [[TMP276]], i64 10
+; OPT-NEXT: [[TMP278:%.*]] = extractelement <128 x i8> [[TMP255]], i32 91
+; OPT-NEXT: [[TMP279:%.*]] = insertelement <16 x i8> [[TMP277]], i8 [[TMP278]], i64 11
+; OPT-NEXT: [[TMP280:%.*]] = extractelement <128 x i8> [[TMP255]], i32 92
+; OPT-NEXT: [[TMP281:%.*]] = insertelement <16 x i8> [[TMP279]], i8 [[TMP280]], i64 12
+; OPT-NEXT: [[TMP282:%.*]] = extractelement <128 x i8> [[TMP255]], i32 93
+; OPT-NEXT: [[TMP283:%.*]] = insertelement <16 x i8> [[TMP281]], i8 [[TMP282]], i64 13
+; OPT-NEXT: [[TMP284:%.*]] = extractelement <128 x i8> [[TMP255]], i32 94
+; OPT-NEXT: [[TMP285:%.*]] = insertelement <16 x i8> [[TMP283]], i8 [[TMP284]], i64 14
+; OPT-NEXT: [[TMP286:%.*]] = extractelement <128 x i8> [[TMP255]], i32 95
+; OPT-NEXT: [[TMP287:%.*]] = insertelement <16 x i8> [[TMP285]], i8 [[TMP286]], i64 15
+; OPT-NEXT: [[SUM:%.*]] = add <16 x i8> [[TMP287]], [[ADD]]
+; OPT-NEXT: store <16 x i8> [[SUM]], ptr addrspace(3) [[OUT]], align 16
+; OPT-NEXT: ret void
+;
+entry:
+ %alloca = alloca [8 x <16 x i8>], align 16, addrspace(5)
+ %gep0 = getelementptr [8 x <16 x i8>], ptr addrspace(5) %alloca, i64 0, i64 0
+ store <16 x i8> %in, ptr addrspace(5) %gep0, align 16
+ %gep1 = getelementptr [8 x <16 x i8>], ptr addrspace(5) %alloca, i64 0, i64 1
+ store <16 x i8> %in, ptr addrspace(5) %gep0, align 16
+ %gep2 = getelementptr [8 x <16 x i8>], ptr addrspace(5) %alloca, i64 0, i64 2
+ store <16 x i8> %in, ptr addrspace(5) %gep0, align 16
+ %gep3 = getelementptr [8 x <16 x i8>], ptr addrspace(5) %alloca, i64 0, i64 3
+ store <16 x i8> %in, ptr addrspace(5) %gep0, align 16
+ %gep4 = getelementptr [8 x <16 x i8>], ptr addrspace(5) %alloca, i64 0, i64 4
+ store <16 x i8> %in, ptr addrspace(5) %gep0, align 16
+ %gep5 = getelementptr [8 x <16 x i8>], ptr addrspace(5) %alloca, i64 0, i64 5
+ store <16 x i8> %in, ptr addrspace(5) %gep0, align 16
+ %gep6 = getelementptr [8 x <16 x i8>], ptr addrspace(5) %alloca, i64 0, i64 6
+ store <16 x i8> %in, ptr addrspace(5) %gep0, align 16
+ %gep7 = getelementptr [8 x <16 x i8>], ptr addrspace(5) %alloca, i64 0, i64 7
+ store <16 x i8> %in, ptr addrspace(5) %gep0, align 16
+ %load = load <16 x i8>, ptr addrspace(5) %gep5, align 16
+ %sum = add <16 x i8> %load, %add
+ store <16 x i8> %sum, ptr addrspace(3) %out, align 16
+ ret void
+}
+
+attributes #0 = {"amdgpu-waves-per-eu"="2,2"}
>From cadbdd65f7fc8cbe3f8cceed1a04a0f365c43f24 Mon Sep 17 00:00:00 2001
From: Kevin Choi <kevin.choi at amd.com>
Date: Mon, 3 Nov 2025 20:42:59 -0600
Subject: [PATCH 2/2] [AMDGPU] NFC, move testcase, only test output of
promote-alloca with vector-combine
---
.../AMDGPU/promote-alloca-array-to-vector.ll | 325 ----------
.../extract-insert-chain-to-shuffles.ll | 567 ++++++++++++++++++
2 files changed, 567 insertions(+), 325 deletions(-)
delete mode 100644 llvm/test/CodeGen/AMDGPU/promote-alloca-array-to-vector.ll
create mode 100644 llvm/test/Transforms/VectorCombine/AMDGPU/extract-insert-chain-to-shuffles.ll
diff --git a/llvm/test/CodeGen/AMDGPU/promote-alloca-array-to-vector.ll b/llvm/test/CodeGen/AMDGPU/promote-alloca-array-to-vector.ll
deleted file mode 100644
index 05a0e39d4a715..0000000000000
--- a/llvm/test/CodeGen/AMDGPU/promote-alloca-array-to-vector.ll
+++ /dev/null
@@ -1,325 +0,0 @@
-; NOTE: Assertions have been autogenerated by utils/update_test_checks.py UTC_ARGS: --version 6
-; RUN: opt -S -mtriple=amdgcn-- -mcpu=gfx1100 -passes=amdgpu-promote-alloca < %s | FileCheck -check-prefix=OPT %s
-
-define amdgpu_kernel void @large_array_vectors_small_users(<16 x i8> %in, <16 x i8> %add, ptr addrspace(3) %out) #0 {
-; OPT-LABEL: define amdgpu_kernel void @large_array_vectors_small_users(
-; OPT-SAME: <16 x i8> [[IN:%.*]], <16 x i8> [[ADD:%.*]], ptr addrspace(3) [[OUT:%.*]]) #[[ATTR0:[0-9]+]] {
-; OPT-NEXT: [[ENTRY:.*:]]
-; OPT-NEXT: [[ALLOCA:%.*]] = freeze <128 x i8> poison
-; OPT-NEXT: [[TMP0:%.*]] = extractelement <16 x i8> [[IN]], i64 0
-; OPT-NEXT: [[TMP1:%.*]] = insertelement <128 x i8> [[ALLOCA]], i8 [[TMP0]], i32 0
-; OPT-NEXT: [[TMP2:%.*]] = extractelement <16 x i8> [[IN]], i64 1
-; OPT-NEXT: [[TMP3:%.*]] = insertelement <128 x i8> [[TMP1]], i8 [[TMP2]], i32 1
-; OPT-NEXT: [[TMP4:%.*]] = extractelement <16 x i8> [[IN]], i64 2
-; OPT-NEXT: [[TMP5:%.*]] = insertelement <128 x i8> [[TMP3]], i8 [[TMP4]], i32 2
-; OPT-NEXT: [[TMP6:%.*]] = extractelement <16 x i8> [[IN]], i64 3
-; OPT-NEXT: [[TMP7:%.*]] = insertelement <128 x i8> [[TMP5]], i8 [[TMP6]], i32 3
-; OPT-NEXT: [[TMP8:%.*]] = extractelement <16 x i8> [[IN]], i64 4
-; OPT-NEXT: [[TMP9:%.*]] = insertelement <128 x i8> [[TMP7]], i8 [[TMP8]], i32 4
-; OPT-NEXT: [[TMP10:%.*]] = extractelement <16 x i8> [[IN]], i64 5
-; OPT-NEXT: [[TMP11:%.*]] = insertelement <128 x i8> [[TMP9]], i8 [[TMP10]], i32 5
-; OPT-NEXT: [[TMP12:%.*]] = extractelement <16 x i8> [[IN]], i64 6
-; OPT-NEXT: [[TMP13:%.*]] = insertelement <128 x i8> [[TMP11]], i8 [[TMP12]], i32 6
-; OPT-NEXT: [[TMP14:%.*]] = extractelement <16 x i8> [[IN]], i64 7
-; OPT-NEXT: [[TMP15:%.*]] = insertelement <128 x i8> [[TMP13]], i8 [[TMP14]], i32 7
-; OPT-NEXT: [[TMP16:%.*]] = extractelement <16 x i8> [[IN]], i64 8
-; OPT-NEXT: [[TMP17:%.*]] = insertelement <128 x i8> [[TMP15]], i8 [[TMP16]], i32 8
-; OPT-NEXT: [[TMP18:%.*]] = extractelement <16 x i8> [[IN]], i64 9
-; OPT-NEXT: [[TMP19:%.*]] = insertelement <128 x i8> [[TMP17]], i8 [[TMP18]], i32 9
-; OPT-NEXT: [[TMP20:%.*]] = extractelement <16 x i8> [[IN]], i64 10
-; OPT-NEXT: [[TMP21:%.*]] = insertelement <128 x i8> [[TMP19]], i8 [[TMP20]], i32 10
-; OPT-NEXT: [[TMP22:%.*]] = extractelement <16 x i8> [[IN]], i64 11
-; OPT-NEXT: [[TMP23:%.*]] = insertelement <128 x i8> [[TMP21]], i8 [[TMP22]], i32 11
-; OPT-NEXT: [[TMP24:%.*]] = extractelement <16 x i8> [[IN]], i64 12
-; OPT-NEXT: [[TMP25:%.*]] = insertelement <128 x i8> [[TMP23]], i8 [[TMP24]], i32 12
-; OPT-NEXT: [[TMP26:%.*]] = extractelement <16 x i8> [[IN]], i64 13
-; OPT-NEXT: [[TMP27:%.*]] = insertelement <128 x i8> [[TMP25]], i8 [[TMP26]], i32 13
-; OPT-NEXT: [[TMP28:%.*]] = extractelement <16 x i8> [[IN]], i64 14
-; OPT-NEXT: [[TMP29:%.*]] = insertelement <128 x i8> [[TMP27]], i8 [[TMP28]], i32 14
-; OPT-NEXT: [[TMP30:%.*]] = extractelement <16 x i8> [[IN]], i64 15
-; OPT-NEXT: [[TMP31:%.*]] = insertelement <128 x i8> [[TMP29]], i8 [[TMP30]], i32 15
-; OPT-NEXT: [[TMP32:%.*]] = extractelement <16 x i8> [[IN]], i64 0
-; OPT-NEXT: [[TMP33:%.*]] = insertelement <128 x i8> [[TMP31]], i8 [[TMP32]], i32 0
-; OPT-NEXT: [[TMP34:%.*]] = extractelement <16 x i8> [[IN]], i64 1
-; OPT-NEXT: [[TMP35:%.*]] = insertelement <128 x i8> [[TMP33]], i8 [[TMP34]], i32 1
-; OPT-NEXT: [[TMP36:%.*]] = extractelement <16 x i8> [[IN]], i64 2
-; OPT-NEXT: [[TMP37:%.*]] = insertelement <128 x i8> [[TMP35]], i8 [[TMP36]], i32 2
-; OPT-NEXT: [[TMP38:%.*]] = extractelement <16 x i8> [[IN]], i64 3
-; OPT-NEXT: [[TMP39:%.*]] = insertelement <128 x i8> [[TMP37]], i8 [[TMP38]], i32 3
-; OPT-NEXT: [[TMP40:%.*]] = extractelement <16 x i8> [[IN]], i64 4
-; OPT-NEXT: [[TMP41:%.*]] = insertelement <128 x i8> [[TMP39]], i8 [[TMP40]], i32 4
-; OPT-NEXT: [[TMP42:%.*]] = extractelement <16 x i8> [[IN]], i64 5
-; OPT-NEXT: [[TMP43:%.*]] = insertelement <128 x i8> [[TMP41]], i8 [[TMP42]], i32 5
-; OPT-NEXT: [[TMP44:%.*]] = extractelement <16 x i8> [[IN]], i64 6
-; OPT-NEXT: [[TMP45:%.*]] = insertelement <128 x i8> [[TMP43]], i8 [[TMP44]], i32 6
-; OPT-NEXT: [[TMP46:%.*]] = extractelement <16 x i8> [[IN]], i64 7
-; OPT-NEXT: [[TMP47:%.*]] = insertelement <128 x i8> [[TMP45]], i8 [[TMP46]], i32 7
-; OPT-NEXT: [[TMP48:%.*]] = extractelement <16 x i8> [[IN]], i64 8
-; OPT-NEXT: [[TMP49:%.*]] = insertelement <128 x i8> [[TMP47]], i8 [[TMP48]], i32 8
-; OPT-NEXT: [[TMP50:%.*]] = extractelement <16 x i8> [[IN]], i64 9
-; OPT-NEXT: [[TMP51:%.*]] = insertelement <128 x i8> [[TMP49]], i8 [[TMP50]], i32 9
-; OPT-NEXT: [[TMP52:%.*]] = extractelement <16 x i8> [[IN]], i64 10
-; OPT-NEXT: [[TMP53:%.*]] = insertelement <128 x i8> [[TMP51]], i8 [[TMP52]], i32 10
-; OPT-NEXT: [[TMP54:%.*]] = extractelement <16 x i8> [[IN]], i64 11
-; OPT-NEXT: [[TMP55:%.*]] = insertelement <128 x i8> [[TMP53]], i8 [[TMP54]], i32 11
-; OPT-NEXT: [[TMP56:%.*]] = extractelement <16 x i8> [[IN]], i64 12
-; OPT-NEXT: [[TMP57:%.*]] = insertelement <128 x i8> [[TMP55]], i8 [[TMP56]], i32 12
-; OPT-NEXT: [[TMP58:%.*]] = extractelement <16 x i8> [[IN]], i64 13
-; OPT-NEXT: [[TMP59:%.*]] = insertelement <128 x i8> [[TMP57]], i8 [[TMP58]], i32 13
-; OPT-NEXT: [[TMP60:%.*]] = extractelement <16 x i8> [[IN]], i64 14
-; OPT-NEXT: [[TMP61:%.*]] = insertelement <128 x i8> [[TMP59]], i8 [[TMP60]], i32 14
-; OPT-NEXT: [[TMP62:%.*]] = extractelement <16 x i8> [[IN]], i64 15
-; OPT-NEXT: [[TMP63:%.*]] = insertelement <128 x i8> [[TMP61]], i8 [[TMP62]], i32 15
-; OPT-NEXT: [[TMP64:%.*]] = extractelement <16 x i8> [[IN]], i64 0
-; OPT-NEXT: [[TMP65:%.*]] = insertelement <128 x i8> [[TMP63]], i8 [[TMP64]], i32 0
-; OPT-NEXT: [[TMP66:%.*]] = extractelement <16 x i8> [[IN]], i64 1
-; OPT-NEXT: [[TMP67:%.*]] = insertelement <128 x i8> [[TMP65]], i8 [[TMP66]], i32 1
-; OPT-NEXT: [[TMP68:%.*]] = extractelement <16 x i8> [[IN]], i64 2
-; OPT-NEXT: [[TMP69:%.*]] = insertelement <128 x i8> [[TMP67]], i8 [[TMP68]], i32 2
-; OPT-NEXT: [[TMP70:%.*]] = extractelement <16 x i8> [[IN]], i64 3
-; OPT-NEXT: [[TMP71:%.*]] = insertelement <128 x i8> [[TMP69]], i8 [[TMP70]], i32 3
-; OPT-NEXT: [[TMP72:%.*]] = extractelement <16 x i8> [[IN]], i64 4
-; OPT-NEXT: [[TMP73:%.*]] = insertelement <128 x i8> [[TMP71]], i8 [[TMP72]], i32 4
-; OPT-NEXT: [[TMP74:%.*]] = extractelement <16 x i8> [[IN]], i64 5
-; OPT-NEXT: [[TMP75:%.*]] = insertelement <128 x i8> [[TMP73]], i8 [[TMP74]], i32 5
-; OPT-NEXT: [[TMP76:%.*]] = extractelement <16 x i8> [[IN]], i64 6
-; OPT-NEXT: [[TMP77:%.*]] = insertelement <128 x i8> [[TMP75]], i8 [[TMP76]], i32 6
-; OPT-NEXT: [[TMP78:%.*]] = extractelement <16 x i8> [[IN]], i64 7
-; OPT-NEXT: [[TMP79:%.*]] = insertelement <128 x i8> [[TMP77]], i8 [[TMP78]], i32 7
-; OPT-NEXT: [[TMP80:%.*]] = extractelement <16 x i8> [[IN]], i64 8
-; OPT-NEXT: [[TMP81:%.*]] = insertelement <128 x i8> [[TMP79]], i8 [[TMP80]], i32 8
-; OPT-NEXT: [[TMP82:%.*]] = extractelement <16 x i8> [[IN]], i64 9
-; OPT-NEXT: [[TMP83:%.*]] = insertelement <128 x i8> [[TMP81]], i8 [[TMP82]], i32 9
-; OPT-NEXT: [[TMP84:%.*]] = extractelement <16 x i8> [[IN]], i64 10
-; OPT-NEXT: [[TMP85:%.*]] = insertelement <128 x i8> [[TMP83]], i8 [[TMP84]], i32 10
-; OPT-NEXT: [[TMP86:%.*]] = extractelement <16 x i8> [[IN]], i64 11
-; OPT-NEXT: [[TMP87:%.*]] = insertelement <128 x i8> [[TMP85]], i8 [[TMP86]], i32 11
-; OPT-NEXT: [[TMP88:%.*]] = extractelement <16 x i8> [[IN]], i64 12
-; OPT-NEXT: [[TMP89:%.*]] = insertelement <128 x i8> [[TMP87]], i8 [[TMP88]], i32 12
-; OPT-NEXT: [[TMP90:%.*]] = extractelement <16 x i8> [[IN]], i64 13
-; OPT-NEXT: [[TMP91:%.*]] = insertelement <128 x i8> [[TMP89]], i8 [[TMP90]], i32 13
-; OPT-NEXT: [[TMP92:%.*]] = extractelement <16 x i8> [[IN]], i64 14
-; OPT-NEXT: [[TMP93:%.*]] = insertelement <128 x i8> [[TMP91]], i8 [[TMP92]], i32 14
-; OPT-NEXT: [[TMP94:%.*]] = extractelement <16 x i8> [[IN]], i64 15
-; OPT-NEXT: [[TMP95:%.*]] = insertelement <128 x i8> [[TMP93]], i8 [[TMP94]], i32 15
-; OPT-NEXT: [[TMP96:%.*]] = extractelement <16 x i8> [[IN]], i64 0
-; OPT-NEXT: [[TMP97:%.*]] = insertelement <128 x i8> [[TMP95]], i8 [[TMP96]], i32 0
-; OPT-NEXT: [[TMP98:%.*]] = extractelement <16 x i8> [[IN]], i64 1
-; OPT-NEXT: [[TMP99:%.*]] = insertelement <128 x i8> [[TMP97]], i8 [[TMP98]], i32 1
-; OPT-NEXT: [[TMP100:%.*]] = extractelement <16 x i8> [[IN]], i64 2
-; OPT-NEXT: [[TMP101:%.*]] = insertelement <128 x i8> [[TMP99]], i8 [[TMP100]], i32 2
-; OPT-NEXT: [[TMP102:%.*]] = extractelement <16 x i8> [[IN]], i64 3
-; OPT-NEXT: [[TMP103:%.*]] = insertelement <128 x i8> [[TMP101]], i8 [[TMP102]], i32 3
-; OPT-NEXT: [[TMP104:%.*]] = extractelement <16 x i8> [[IN]], i64 4
-; OPT-NEXT: [[TMP105:%.*]] = insertelement <128 x i8> [[TMP103]], i8 [[TMP104]], i32 4
-; OPT-NEXT: [[TMP106:%.*]] = extractelement <16 x i8> [[IN]], i64 5
-; OPT-NEXT: [[TMP107:%.*]] = insertelement <128 x i8> [[TMP105]], i8 [[TMP106]], i32 5
-; OPT-NEXT: [[TMP108:%.*]] = extractelement <16 x i8> [[IN]], i64 6
-; OPT-NEXT: [[TMP109:%.*]] = insertelement <128 x i8> [[TMP107]], i8 [[TMP108]], i32 6
-; OPT-NEXT: [[TMP110:%.*]] = extractelement <16 x i8> [[IN]], i64 7
-; OPT-NEXT: [[TMP111:%.*]] = insertelement <128 x i8> [[TMP109]], i8 [[TMP110]], i32 7
-; OPT-NEXT: [[TMP112:%.*]] = extractelement <16 x i8> [[IN]], i64 8
-; OPT-NEXT: [[TMP113:%.*]] = insertelement <128 x i8> [[TMP111]], i8 [[TMP112]], i32 8
-; OPT-NEXT: [[TMP114:%.*]] = extractelement <16 x i8> [[IN]], i64 9
-; OPT-NEXT: [[TMP115:%.*]] = insertelement <128 x i8> [[TMP113]], i8 [[TMP114]], i32 9
-; OPT-NEXT: [[TMP116:%.*]] = extractelement <16 x i8> [[IN]], i64 10
-; OPT-NEXT: [[TMP117:%.*]] = insertelement <128 x i8> [[TMP115]], i8 [[TMP116]], i32 10
-; OPT-NEXT: [[TMP118:%.*]] = extractelement <16 x i8> [[IN]], i64 11
-; OPT-NEXT: [[TMP119:%.*]] = insertelement <128 x i8> [[TMP117]], i8 [[TMP118]], i32 11
-; OPT-NEXT: [[TMP120:%.*]] = extractelement <16 x i8> [[IN]], i64 12
-; OPT-NEXT: [[TMP121:%.*]] = insertelement <128 x i8> [[TMP119]], i8 [[TMP120]], i32 12
-; OPT-NEXT: [[TMP122:%.*]] = extractelement <16 x i8> [[IN]], i64 13
-; OPT-NEXT: [[TMP123:%.*]] = insertelement <128 x i8> [[TMP121]], i8 [[TMP122]], i32 13
-; OPT-NEXT: [[TMP124:%.*]] = extractelement <16 x i8> [[IN]], i64 14
-; OPT-NEXT: [[TMP125:%.*]] = insertelement <128 x i8> [[TMP123]], i8 [[TMP124]], i32 14
-; OPT-NEXT: [[TMP126:%.*]] = extractelement <16 x i8> [[IN]], i64 15
-; OPT-NEXT: [[TMP127:%.*]] = insertelement <128 x i8> [[TMP125]], i8 [[TMP126]], i32 15
-; OPT-NEXT: [[TMP128:%.*]] = extractelement <16 x i8> [[IN]], i64 0
-; OPT-NEXT: [[TMP129:%.*]] = insertelement <128 x i8> [[TMP127]], i8 [[TMP128]], i32 0
-; OPT-NEXT: [[TMP130:%.*]] = extractelement <16 x i8> [[IN]], i64 1
-; OPT-NEXT: [[TMP131:%.*]] = insertelement <128 x i8> [[TMP129]], i8 [[TMP130]], i32 1
-; OPT-NEXT: [[TMP132:%.*]] = extractelement <16 x i8> [[IN]], i64 2
-; OPT-NEXT: [[TMP133:%.*]] = insertelement <128 x i8> [[TMP131]], i8 [[TMP132]], i32 2
-; OPT-NEXT: [[TMP134:%.*]] = extractelement <16 x i8> [[IN]], i64 3
-; OPT-NEXT: [[TMP135:%.*]] = insertelement <128 x i8> [[TMP133]], i8 [[TMP134]], i32 3
-; OPT-NEXT: [[TMP136:%.*]] = extractelement <16 x i8> [[IN]], i64 4
-; OPT-NEXT: [[TMP137:%.*]] = insertelement <128 x i8> [[TMP135]], i8 [[TMP136]], i32 4
-; OPT-NEXT: [[TMP138:%.*]] = extractelement <16 x i8> [[IN]], i64 5
-; OPT-NEXT: [[TMP139:%.*]] = insertelement <128 x i8> [[TMP137]], i8 [[TMP138]], i32 5
-; OPT-NEXT: [[TMP140:%.*]] = extractelement <16 x i8> [[IN]], i64 6
-; OPT-NEXT: [[TMP141:%.*]] = insertelement <128 x i8> [[TMP139]], i8 [[TMP140]], i32 6
-; OPT-NEXT: [[TMP142:%.*]] = extractelement <16 x i8> [[IN]], i64 7
-; OPT-NEXT: [[TMP143:%.*]] = insertelement <128 x i8> [[TMP141]], i8 [[TMP142]], i32 7
-; OPT-NEXT: [[TMP144:%.*]] = extractelement <16 x i8> [[IN]], i64 8
-; OPT-NEXT: [[TMP145:%.*]] = insertelement <128 x i8> [[TMP143]], i8 [[TMP144]], i32 8
-; OPT-NEXT: [[TMP146:%.*]] = extractelement <16 x i8> [[IN]], i64 9
-; OPT-NEXT: [[TMP147:%.*]] = insertelement <128 x i8> [[TMP145]], i8 [[TMP146]], i32 9
-; OPT-NEXT: [[TMP148:%.*]] = extractelement <16 x i8> [[IN]], i64 10
-; OPT-NEXT: [[TMP149:%.*]] = insertelement <128 x i8> [[TMP147]], i8 [[TMP148]], i32 10
-; OPT-NEXT: [[TMP150:%.*]] = extractelement <16 x i8> [[IN]], i64 11
-; OPT-NEXT: [[TMP151:%.*]] = insertelement <128 x i8> [[TMP149]], i8 [[TMP150]], i32 11
-; OPT-NEXT: [[TMP152:%.*]] = extractelement <16 x i8> [[IN]], i64 12
-; OPT-NEXT: [[TMP153:%.*]] = insertelement <128 x i8> [[TMP151]], i8 [[TMP152]], i32 12
-; OPT-NEXT: [[TMP154:%.*]] = extractelement <16 x i8> [[IN]], i64 13
-; OPT-NEXT: [[TMP155:%.*]] = insertelement <128 x i8> [[TMP153]], i8 [[TMP154]], i32 13
-; OPT-NEXT: [[TMP156:%.*]] = extractelement <16 x i8> [[IN]], i64 14
-; OPT-NEXT: [[TMP157:%.*]] = insertelement <128 x i8> [[TMP155]], i8 [[TMP156]], i32 14
-; OPT-NEXT: [[TMP158:%.*]] = extractelement <16 x i8> [[IN]], i64 15
-; OPT-NEXT: [[TMP159:%.*]] = insertelement <128 x i8> [[TMP157]], i8 [[TMP158]], i32 15
-; OPT-NEXT: [[TMP160:%.*]] = extractelement <16 x i8> [[IN]], i64 0
-; OPT-NEXT: [[TMP161:%.*]] = insertelement <128 x i8> [[TMP159]], i8 [[TMP160]], i32 0
-; OPT-NEXT: [[TMP162:%.*]] = extractelement <16 x i8> [[IN]], i64 1
-; OPT-NEXT: [[TMP163:%.*]] = insertelement <128 x i8> [[TMP161]], i8 [[TMP162]], i32 1
-; OPT-NEXT: [[TMP164:%.*]] = extractelement <16 x i8> [[IN]], i64 2
-; OPT-NEXT: [[TMP165:%.*]] = insertelement <128 x i8> [[TMP163]], i8 [[TMP164]], i32 2
-; OPT-NEXT: [[TMP166:%.*]] = extractelement <16 x i8> [[IN]], i64 3
-; OPT-NEXT: [[TMP167:%.*]] = insertelement <128 x i8> [[TMP165]], i8 [[TMP166]], i32 3
-; OPT-NEXT: [[TMP168:%.*]] = extractelement <16 x i8> [[IN]], i64 4
-; OPT-NEXT: [[TMP169:%.*]] = insertelement <128 x i8> [[TMP167]], i8 [[TMP168]], i32 4
-; OPT-NEXT: [[TMP170:%.*]] = extractelement <16 x i8> [[IN]], i64 5
-; OPT-NEXT: [[TMP171:%.*]] = insertelement <128 x i8> [[TMP169]], i8 [[TMP170]], i32 5
-; OPT-NEXT: [[TMP172:%.*]] = extractelement <16 x i8> [[IN]], i64 6
-; OPT-NEXT: [[TMP173:%.*]] = insertelement <128 x i8> [[TMP171]], i8 [[TMP172]], i32 6
-; OPT-NEXT: [[TMP174:%.*]] = extractelement <16 x i8> [[IN]], i64 7
-; OPT-NEXT: [[TMP175:%.*]] = insertelement <128 x i8> [[TMP173]], i8 [[TMP174]], i32 7
-; OPT-NEXT: [[TMP176:%.*]] = extractelement <16 x i8> [[IN]], i64 8
-; OPT-NEXT: [[TMP177:%.*]] = insertelement <128 x i8> [[TMP175]], i8 [[TMP176]], i32 8
-; OPT-NEXT: [[TMP178:%.*]] = extractelement <16 x i8> [[IN]], i64 9
-; OPT-NEXT: [[TMP179:%.*]] = insertelement <128 x i8> [[TMP177]], i8 [[TMP178]], i32 9
-; OPT-NEXT: [[TMP180:%.*]] = extractelement <16 x i8> [[IN]], i64 10
-; OPT-NEXT: [[TMP181:%.*]] = insertelement <128 x i8> [[TMP179]], i8 [[TMP180]], i32 10
-; OPT-NEXT: [[TMP182:%.*]] = extractelement <16 x i8> [[IN]], i64 11
-; OPT-NEXT: [[TMP183:%.*]] = insertelement <128 x i8> [[TMP181]], i8 [[TMP182]], i32 11
-; OPT-NEXT: [[TMP184:%.*]] = extractelement <16 x i8> [[IN]], i64 12
-; OPT-NEXT: [[TMP185:%.*]] = insertelement <128 x i8> [[TMP183]], i8 [[TMP184]], i32 12
-; OPT-NEXT: [[TMP186:%.*]] = extractelement <16 x i8> [[IN]], i64 13
-; OPT-NEXT: [[TMP187:%.*]] = insertelement <128 x i8> [[TMP185]], i8 [[TMP186]], i32 13
-; OPT-NEXT: [[TMP188:%.*]] = extractelement <16 x i8> [[IN]], i64 14
-; OPT-NEXT: [[TMP189:%.*]] = insertelement <128 x i8> [[TMP187]], i8 [[TMP188]], i32 14
-; OPT-NEXT: [[TMP190:%.*]] = extractelement <16 x i8> [[IN]], i64 15
-; OPT-NEXT: [[TMP191:%.*]] = insertelement <128 x i8> [[TMP189]], i8 [[TMP190]], i32 15
-; OPT-NEXT: [[TMP192:%.*]] = extractelement <16 x i8> [[IN]], i64 0
-; OPT-NEXT: [[TMP193:%.*]] = insertelement <128 x i8> [[TMP191]], i8 [[TMP192]], i32 0
-; OPT-NEXT: [[TMP194:%.*]] = extractelement <16 x i8> [[IN]], i64 1
-; OPT-NEXT: [[TMP195:%.*]] = insertelement <128 x i8> [[TMP193]], i8 [[TMP194]], i32 1
-; OPT-NEXT: [[TMP196:%.*]] = extractelement <16 x i8> [[IN]], i64 2
-; OPT-NEXT: [[TMP197:%.*]] = insertelement <128 x i8> [[TMP195]], i8 [[TMP196]], i32 2
-; OPT-NEXT: [[TMP198:%.*]] = extractelement <16 x i8> [[IN]], i64 3
-; OPT-NEXT: [[TMP199:%.*]] = insertelement <128 x i8> [[TMP197]], i8 [[TMP198]], i32 3
-; OPT-NEXT: [[TMP200:%.*]] = extractelement <16 x i8> [[IN]], i64 4
-; OPT-NEXT: [[TMP201:%.*]] = insertelement <128 x i8> [[TMP199]], i8 [[TMP200]], i32 4
-; OPT-NEXT: [[TMP202:%.*]] = extractelement <16 x i8> [[IN]], i64 5
-; OPT-NEXT: [[TMP203:%.*]] = insertelement <128 x i8> [[TMP201]], i8 [[TMP202]], i32 5
-; OPT-NEXT: [[TMP204:%.*]] = extractelement <16 x i8> [[IN]], i64 6
-; OPT-NEXT: [[TMP205:%.*]] = insertelement <128 x i8> [[TMP203]], i8 [[TMP204]], i32 6
-; OPT-NEXT: [[TMP206:%.*]] = extractelement <16 x i8> [[IN]], i64 7
-; OPT-NEXT: [[TMP207:%.*]] = insertelement <128 x i8> [[TMP205]], i8 [[TMP206]], i32 7
-; OPT-NEXT: [[TMP208:%.*]] = extractelement <16 x i8> [[IN]], i64 8
-; OPT-NEXT: [[TMP209:%.*]] = insertelement <128 x i8> [[TMP207]], i8 [[TMP208]], i32 8
-; OPT-NEXT: [[TMP210:%.*]] = extractelement <16 x i8> [[IN]], i64 9
-; OPT-NEXT: [[TMP211:%.*]] = insertelement <128 x i8> [[TMP209]], i8 [[TMP210]], i32 9
-; OPT-NEXT: [[TMP212:%.*]] = extractelement <16 x i8> [[IN]], i64 10
-; OPT-NEXT: [[TMP213:%.*]] = insertelement <128 x i8> [[TMP211]], i8 [[TMP212]], i32 10
-; OPT-NEXT: [[TMP214:%.*]] = extractelement <16 x i8> [[IN]], i64 11
-; OPT-NEXT: [[TMP215:%.*]] = insertelement <128 x i8> [[TMP213]], i8 [[TMP214]], i32 11
-; OPT-NEXT: [[TMP216:%.*]] = extractelement <16 x i8> [[IN]], i64 12
-; OPT-NEXT: [[TMP217:%.*]] = insertelement <128 x i8> [[TMP215]], i8 [[TMP216]], i32 12
-; OPT-NEXT: [[TMP218:%.*]] = extractelement <16 x i8> [[IN]], i64 13
-; OPT-NEXT: [[TMP219:%.*]] = insertelement <128 x i8> [[TMP217]], i8 [[TMP218]], i32 13
-; OPT-NEXT: [[TMP220:%.*]] = extractelement <16 x i8> [[IN]], i64 14
-; OPT-NEXT: [[TMP221:%.*]] = insertelement <128 x i8> [[TMP219]], i8 [[TMP220]], i32 14
-; OPT-NEXT: [[TMP222:%.*]] = extractelement <16 x i8> [[IN]], i64 15
-; OPT-NEXT: [[TMP223:%.*]] = insertelement <128 x i8> [[TMP221]], i8 [[TMP222]], i32 15
-; OPT-NEXT: [[TMP224:%.*]] = extractelement <16 x i8> [[IN]], i64 0
-; OPT-NEXT: [[TMP225:%.*]] = insertelement <128 x i8> [[TMP223]], i8 [[TMP224]], i32 0
-; OPT-NEXT: [[TMP226:%.*]] = extractelement <16 x i8> [[IN]], i64 1
-; OPT-NEXT: [[TMP227:%.*]] = insertelement <128 x i8> [[TMP225]], i8 [[TMP226]], i32 1
-; OPT-NEXT: [[TMP228:%.*]] = extractelement <16 x i8> [[IN]], i64 2
-; OPT-NEXT: [[TMP229:%.*]] = insertelement <128 x i8> [[TMP227]], i8 [[TMP228]], i32 2
-; OPT-NEXT: [[TMP230:%.*]] = extractelement <16 x i8> [[IN]], i64 3
-; OPT-NEXT: [[TMP231:%.*]] = insertelement <128 x i8> [[TMP229]], i8 [[TMP230]], i32 3
-; OPT-NEXT: [[TMP232:%.*]] = extractelement <16 x i8> [[IN]], i64 4
-; OPT-NEXT: [[TMP233:%.*]] = insertelement <128 x i8> [[TMP231]], i8 [[TMP232]], i32 4
-; OPT-NEXT: [[TMP234:%.*]] = extractelement <16 x i8> [[IN]], i64 5
-; OPT-NEXT: [[TMP235:%.*]] = insertelement <128 x i8> [[TMP233]], i8 [[TMP234]], i32 5
-; OPT-NEXT: [[TMP236:%.*]] = extractelement <16 x i8> [[IN]], i64 6
-; OPT-NEXT: [[TMP237:%.*]] = insertelement <128 x i8> [[TMP235]], i8 [[TMP236]], i32 6
-; OPT-NEXT: [[TMP238:%.*]] = extractelement <16 x i8> [[IN]], i64 7
-; OPT-NEXT: [[TMP239:%.*]] = insertelement <128 x i8> [[TMP237]], i8 [[TMP238]], i32 7
-; OPT-NEXT: [[TMP240:%.*]] = extractelement <16 x i8> [[IN]], i64 8
-; OPT-NEXT: [[TMP241:%.*]] = insertelement <128 x i8> [[TMP239]], i8 [[TMP240]], i32 8
-; OPT-NEXT: [[TMP242:%.*]] = extractelement <16 x i8> [[IN]], i64 9
-; OPT-NEXT: [[TMP243:%.*]] = insertelement <128 x i8> [[TMP241]], i8 [[TMP242]], i32 9
-; OPT-NEXT: [[TMP244:%.*]] = extractelement <16 x i8> [[IN]], i64 10
-; OPT-NEXT: [[TMP245:%.*]] = insertelement <128 x i8> [[TMP243]], i8 [[TMP244]], i32 10
-; OPT-NEXT: [[TMP246:%.*]] = extractelement <16 x i8> [[IN]], i64 11
-; OPT-NEXT: [[TMP247:%.*]] = insertelement <128 x i8> [[TMP245]], i8 [[TMP246]], i32 11
-; OPT-NEXT: [[TMP248:%.*]] = extractelement <16 x i8> [[IN]], i64 12
-; OPT-NEXT: [[TMP249:%.*]] = insertelement <128 x i8> [[TMP247]], i8 [[TMP248]], i32 12
-; OPT-NEXT: [[TMP250:%.*]] = extractelement <16 x i8> [[IN]], i64 13
-; OPT-NEXT: [[TMP251:%.*]] = insertelement <128 x i8> [[TMP249]], i8 [[TMP250]], i32 13
-; OPT-NEXT: [[TMP252:%.*]] = extractelement <16 x i8> [[IN]], i64 14
-; OPT-NEXT: [[TMP253:%.*]] = insertelement <128 x i8> [[TMP251]], i8 [[TMP252]], i32 14
-; OPT-NEXT: [[TMP254:%.*]] = extractelement <16 x i8> [[IN]], i64 15
-; OPT-NEXT: [[TMP255:%.*]] = insertelement <128 x i8> [[TMP253]], i8 [[TMP254]], i32 15
-; OPT-NEXT: [[TMP256:%.*]] = extractelement <128 x i8> [[TMP255]], i32 80
-; OPT-NEXT: [[TMP257:%.*]] = insertelement <16 x i8> poison, i8 [[TMP256]], i64 0
-; OPT-NEXT: [[TMP258:%.*]] = extractelement <128 x i8> [[TMP255]], i32 81
-; OPT-NEXT: [[TMP259:%.*]] = insertelement <16 x i8> [[TMP257]], i8 [[TMP258]], i64 1
-; OPT-NEXT: [[TMP260:%.*]] = extractelement <128 x i8> [[TMP255]], i32 82
-; OPT-NEXT: [[TMP261:%.*]] = insertelement <16 x i8> [[TMP259]], i8 [[TMP260]], i64 2
-; OPT-NEXT: [[TMP262:%.*]] = extractelement <128 x i8> [[TMP255]], i32 83
-; OPT-NEXT: [[TMP263:%.*]] = insertelement <16 x i8> [[TMP261]], i8 [[TMP262]], i64 3
-; OPT-NEXT: [[TMP264:%.*]] = extractelement <128 x i8> [[TMP255]], i32 84
-; OPT-NEXT: [[TMP265:%.*]] = insertelement <16 x i8> [[TMP263]], i8 [[TMP264]], i64 4
-; OPT-NEXT: [[TMP266:%.*]] = extractelement <128 x i8> [[TMP255]], i32 85
-; OPT-NEXT: [[TMP267:%.*]] = insertelement <16 x i8> [[TMP265]], i8 [[TMP266]], i64 5
-; OPT-NEXT: [[TMP268:%.*]] = extractelement <128 x i8> [[TMP255]], i32 86
-; OPT-NEXT: [[TMP269:%.*]] = insertelement <16 x i8> [[TMP267]], i8 [[TMP268]], i64 6
-; OPT-NEXT: [[TMP270:%.*]] = extractelement <128 x i8> [[TMP255]], i32 87
-; OPT-NEXT: [[TMP271:%.*]] = insertelement <16 x i8> [[TMP269]], i8 [[TMP270]], i64 7
-; OPT-NEXT: [[TMP272:%.*]] = extractelement <128 x i8> [[TMP255]], i32 88
-; OPT-NEXT: [[TMP273:%.*]] = insertelement <16 x i8> [[TMP271]], i8 [[TMP272]], i64 8
-; OPT-NEXT: [[TMP274:%.*]] = extractelement <128 x i8> [[TMP255]], i32 89
-; OPT-NEXT: [[TMP275:%.*]] = insertelement <16 x i8> [[TMP273]], i8 [[TMP274]], i64 9
-; OPT-NEXT: [[TMP276:%.*]] = extractelement <128 x i8> [[TMP255]], i32 90
-; OPT-NEXT: [[TMP277:%.*]] = insertelement <16 x i8> [[TMP275]], i8 [[TMP276]], i64 10
-; OPT-NEXT: [[TMP278:%.*]] = extractelement <128 x i8> [[TMP255]], i32 91
-; OPT-NEXT: [[TMP279:%.*]] = insertelement <16 x i8> [[TMP277]], i8 [[TMP278]], i64 11
-; OPT-NEXT: [[TMP280:%.*]] = extractelement <128 x i8> [[TMP255]], i32 92
-; OPT-NEXT: [[TMP281:%.*]] = insertelement <16 x i8> [[TMP279]], i8 [[TMP280]], i64 12
-; OPT-NEXT: [[TMP282:%.*]] = extractelement <128 x i8> [[TMP255]], i32 93
-; OPT-NEXT: [[TMP283:%.*]] = insertelement <16 x i8> [[TMP281]], i8 [[TMP282]], i64 13
-; OPT-NEXT: [[TMP284:%.*]] = extractelement <128 x i8> [[TMP255]], i32 94
-; OPT-NEXT: [[TMP285:%.*]] = insertelement <16 x i8> [[TMP283]], i8 [[TMP284]], i64 14
-; OPT-NEXT: [[TMP286:%.*]] = extractelement <128 x i8> [[TMP255]], i32 95
-; OPT-NEXT: [[TMP287:%.*]] = insertelement <16 x i8> [[TMP285]], i8 [[TMP286]], i64 15
-; OPT-NEXT: [[SUM:%.*]] = add <16 x i8> [[TMP287]], [[ADD]]
-; OPT-NEXT: store <16 x i8> [[SUM]], ptr addrspace(3) [[OUT]], align 16
-; OPT-NEXT: ret void
-;
-entry:
- %alloca = alloca [8 x <16 x i8>], align 16, addrspace(5)
- %gep0 = getelementptr [8 x <16 x i8>], ptr addrspace(5) %alloca, i64 0, i64 0
- store <16 x i8> %in, ptr addrspace(5) %gep0, align 16
- %gep1 = getelementptr [8 x <16 x i8>], ptr addrspace(5) %alloca, i64 0, i64 1
- store <16 x i8> %in, ptr addrspace(5) %gep0, align 16
- %gep2 = getelementptr [8 x <16 x i8>], ptr addrspace(5) %alloca, i64 0, i64 2
- store <16 x i8> %in, ptr addrspace(5) %gep0, align 16
- %gep3 = getelementptr [8 x <16 x i8>], ptr addrspace(5) %alloca, i64 0, i64 3
- store <16 x i8> %in, ptr addrspace(5) %gep0, align 16
- %gep4 = getelementptr [8 x <16 x i8>], ptr addrspace(5) %alloca, i64 0, i64 4
- store <16 x i8> %in, ptr addrspace(5) %gep0, align 16
- %gep5 = getelementptr [8 x <16 x i8>], ptr addrspace(5) %alloca, i64 0, i64 5
- store <16 x i8> %in, ptr addrspace(5) %gep0, align 16
- %gep6 = getelementptr [8 x <16 x i8>], ptr addrspace(5) %alloca, i64 0, i64 6
- store <16 x i8> %in, ptr addrspace(5) %gep0, align 16
- %gep7 = getelementptr [8 x <16 x i8>], ptr addrspace(5) %alloca, i64 0, i64 7
- store <16 x i8> %in, ptr addrspace(5) %gep0, align 16
- %load = load <16 x i8>, ptr addrspace(5) %gep5, align 16
- %sum = add <16 x i8> %load, %add
- store <16 x i8> %sum, ptr addrspace(3) %out, align 16
- ret void
-}
-
-attributes #0 = {"amdgpu-waves-per-eu"="2,2"}
diff --git a/llvm/test/Transforms/VectorCombine/AMDGPU/extract-insert-chain-to-shuffles.ll b/llvm/test/Transforms/VectorCombine/AMDGPU/extract-insert-chain-to-shuffles.ll
new file mode 100644
index 0000000000000..4b551fad5b43a
--- /dev/null
+++ b/llvm/test/Transforms/VectorCombine/AMDGPU/extract-insert-chain-to-shuffles.ll
@@ -0,0 +1,567 @@
+; NOTE: Assertions have been autogenerated by utils/update_test_checks.py UTC_ARGS: --version 6
+; RUN: opt -S -mtriple=amdgcn-- -mcpu=gfx1100 -passes=vector-combine < %s | FileCheck -check-prefix=OPT %s
+
+; Generated from amdgpu-promote-alloca on array of vectors
+; VectorCombiner should recognize chain of extract-insert vectors
+; and turn them into one or two shuffles
+define amdgpu_kernel void @extract_insert_chain_to_shuffles(<16 x i8> %in, <16 x i8> %add, ptr addrspace(3) %out) #0 {
+; OPT-LABEL: define amdgpu_kernel void @extract_insert_chain_to_shuffles(
+; OPT-SAME: <16 x i8> [[IN:%.*]], <16 x i8> [[ADD:%.*]], ptr addrspace(3) [[OUT:%.*]]) #[[ATTR0:[0-9]+]] {
+; OPT-NEXT: [[ENTRY:.*:]]
+; OPT-NEXT: [[ALLOCA:%.*]] = freeze <128 x i8> poison
+; OPT-NEXT: [[TMP0:%.*]] = shufflevector <16 x i8> [[IN]], <16 x i8> poison, <128 x i32> <i32 0, i32 poison, i32 poison, i32 poison, i32 poison, i32 poison, i32 poison, i32 poison, i32 poison, i32 poison, i32 poison, i32 poison, i32 poison, i32 poison, i32 poison, i32 poison, i32 poison, i32 poison, i32 poison, i32 poison, i32 poison, i32 poison, i32 poison, i32 poison, i32 poison, i32 poison, i32 poison, i32 poison, i32 poison, i32 poison, i32 poison, i32 poison, i32 poison, i32 poison, i32 poison, i32 poison, i32 poison, i32 poison, i32 poison, i32 poison, i32 poison, i32 poison, i32 poison, i32 poison, i32 poison, i32 poison, i32 poison, i32 poison, i32 poison, i32 poison, i32 poison, i32 poison, i32 poison, i32 poison, i32 poison, i32 poison, i32 poison, i32 poison, i32 poison, i32 poison, i32 poison, i32 poison, i32 poison, i32 poison, i32 poison, i32 poison, i32 poison, i32 poison, i32 poison, i32 poison, i32 poison, i32 poison, i32 poison, i32 poison, i32 poison, i32 poison, i32 poison, i32 poison, i32 poison, i32 poison, i32 poison, i32 poison, i32 poison, i32 poison, i32 poison, i32 poison, i32 poison, i32 poison, i32 poison, i32 poison, i32 poison, i32 poison, i32 poison, i32 poison, i32 poison, i32 poison, i32 poison, i32 poison, i32 poison, i32 poison, i32 poison, i32 poison, i32 poison, i32 poison, i32 poison, i32 poison, i32 poison, i32 poison, i32 poison, i32 poison, i32 poison, i32 poison, i32 poison, i32 poison, i32 poison, i32 poison, i32 poison, i32 poison, i32 poison, i32 poison, i32 poison, i32 poison, i32 poison, i32 poison, i32 poison, i32 poison, i32 poison, i32 poison>
+; OPT-NEXT: [[TMP1:%.*]] = shufflevector <128 x i8> [[ALLOCA]], <128 x i8> [[TMP0]], <128 x i32> <i32 128, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15, i32 16, i32 17, i32 18, i32 19, i32 20, i32 21, i32 22, i32 23, i32 24, i32 25, i32 26, i32 27, i32 28, i32 29, i32 30, i32 31, i32 32, i32 33, i32 34, i32 35, i32 36, i32 37, i32 38, i32 39, i32 40, i32 41, i32 42, i32 43, i32 44, i32 45, i32 46, i32 47, i32 48, i32 49, i32 50, i32 51, i32 52, i32 53, i32 54, i32 55, i32 56, i32 57, i32 58, i32 59, i32 60, i32 61, i32 62, i32 63, i32 64, i32 65, i32 66, i32 67, i32 68, i32 69, i32 70, i32 71, i32 72, i32 73, i32 74, i32 75, i32 76, i32 77, i32 78, i32 79, i32 80, i32 81, i32 82, i32 83, i32 84, i32 85, i32 86, i32 87, i32 88, i32 89, i32 90, i32 91, i32 92, i32 93, i32 94, i32 95, i32 96, i32 97, i32 98, i32 99, i32 100, i32 101, i32 102, i32 103, i32 104, i32 105, i32 106, i32 107, i32 108, i32 109, i32 110, i32 111, i32 112, i32 113, i32 114, i32 115, i32 116, i32 117, i32 118, i32 119, i32 120, i32 121, i32 122, i32 123, i32 124, i32 125, i32 126, i32 127>
+; OPT-NEXT: [[TMP2:%.*]] = extractelement <16 x i8> [[IN]], i64 1
+; OPT-NEXT: [[TMP3:%.*]] = insertelement <128 x i8> [[TMP1]], i8 [[TMP2]], i32 1
+; OPT-NEXT: [[TMP4:%.*]] = extractelement <16 x i8> [[IN]], i64 2
+; OPT-NEXT: [[TMP5:%.*]] = insertelement <128 x i8> [[TMP3]], i8 [[TMP4]], i32 2
+; OPT-NEXT: [[TMP6:%.*]] = extractelement <16 x i8> [[IN]], i64 3
+; OPT-NEXT: [[TMP7:%.*]] = insertelement <128 x i8> [[TMP5]], i8 [[TMP6]], i32 3
+; OPT-NEXT: [[TMP8:%.*]] = extractelement <16 x i8> [[IN]], i64 4
+; OPT-NEXT: [[TMP9:%.*]] = insertelement <128 x i8> [[TMP7]], i8 [[TMP8]], i32 4
+; OPT-NEXT: [[TMP10:%.*]] = extractelement <16 x i8> [[IN]], i64 5
+; OPT-NEXT: [[TMP11:%.*]] = insertelement <128 x i8> [[TMP9]], i8 [[TMP10]], i32 5
+; OPT-NEXT: [[TMP12:%.*]] = extractelement <16 x i8> [[IN]], i64 6
+; OPT-NEXT: [[TMP13:%.*]] = insertelement <128 x i8> [[TMP11]], i8 [[TMP12]], i32 6
+; OPT-NEXT: [[TMP14:%.*]] = extractelement <16 x i8> [[IN]], i64 7
+; OPT-NEXT: [[TMP15:%.*]] = insertelement <128 x i8> [[TMP13]], i8 [[TMP14]], i32 7
+; OPT-NEXT: [[TMP16:%.*]] = extractelement <16 x i8> [[IN]], i64 8
+; OPT-NEXT: [[TMP17:%.*]] = insertelement <128 x i8> [[TMP15]], i8 [[TMP16]], i32 8
+; OPT-NEXT: [[TMP18:%.*]] = extractelement <16 x i8> [[IN]], i64 9
+; OPT-NEXT: [[TMP19:%.*]] = insertelement <128 x i8> [[TMP17]], i8 [[TMP18]], i32 9
+; OPT-NEXT: [[TMP20:%.*]] = extractelement <16 x i8> [[IN]], i64 10
+; OPT-NEXT: [[TMP21:%.*]] = insertelement <128 x i8> [[TMP19]], i8 [[TMP20]], i32 10
+; OPT-NEXT: [[TMP22:%.*]] = extractelement <16 x i8> [[IN]], i64 11
+; OPT-NEXT: [[TMP23:%.*]] = insertelement <128 x i8> [[TMP21]], i8 [[TMP22]], i32 11
+; OPT-NEXT: [[TMP24:%.*]] = extractelement <16 x i8> [[IN]], i64 12
+; OPT-NEXT: [[TMP25:%.*]] = insertelement <128 x i8> [[TMP23]], i8 [[TMP24]], i32 12
+; OPT-NEXT: [[TMP26:%.*]] = extractelement <16 x i8> [[IN]], i64 13
+; OPT-NEXT: [[TMP27:%.*]] = insertelement <128 x i8> [[TMP25]], i8 [[TMP26]], i32 13
+; OPT-NEXT: [[TMP28:%.*]] = extractelement <16 x i8> [[IN]], i64 14
+; OPT-NEXT: [[TMP29:%.*]] = insertelement <128 x i8> [[TMP27]], i8 [[TMP28]], i32 14
+; OPT-NEXT: [[TMP30:%.*]] = extractelement <16 x i8> [[IN]], i64 15
+; OPT-NEXT: [[TMP31:%.*]] = insertelement <128 x i8> [[TMP29]], i8 [[TMP30]], i32 15
+; OPT-NEXT: [[TMP32:%.*]] = shufflevector <16 x i8> [[IN]], <16 x i8> poison, <128 x i32> <i32 0, i32 poison, i32 poison, i32 poison, i32 poison, i32 poison, i32 poison, i32 poison, i32 poison, i32 poison, i32 poison, i32 poison, i32 poison, i32 poison, i32 poison, i32 poison, i32 poison, i32 poison, i32 poison, i32 poison, i32 poison, i32 poison, i32 poison, i32 poison, i32 poison, i32 poison, i32 poison, i32 poison, i32 poison, i32 poison, i32 poison, i32 poison, i32 poison, i32 poison, i32 poison, i32 poison, i32 poison, i32 poison, i32 poison, i32 poison, i32 poison, i32 poison, i32 poison, i32 poison, i32 poison, i32 poison, i32 poison, i32 poison, i32 poison, i32 poison, i32 poison, i32 poison, i32 poison, i32 poison, i32 poison, i32 poison, i32 poison, i32 poison, i32 poison, i32 poison, i32 poison, i32 poison, i32 poison, i32 poison, i32 poison, i32 poison, i32 poison, i32 poison, i32 poison, i32 poison, i32 poison, i32 poison, i32 poison, i32 poison, i32 poison, i32 poison, i32 poison, i32 poison, i32 poison, i32 poison, i32 poison, i32 poison, i32 poison, i32 poison, i32 poison, i32 poison, i32 poison, i32 poison, i32 poison, i32 poison, i32 poison, i32 poison, i32 poison, i32 poison, i32 poison, i32 poison, i32 poison, i32 poison, i32 poison, i32 poison, i32 poison, i32 poison, i32 poison, i32 poison, i32 poison, i32 poison, i32 poison, i32 poison, i32 poison, i32 poison, i32 poison, i32 poison, i32 poison, i32 poison, i32 poison, i32 poison, i32 poison, i32 poison, i32 poison, i32 poison, i32 poison, i32 poison, i32 poison, i32 poison, i32 poison, i32 poison, i32 poison, i32 poison>
+; OPT-NEXT: [[TMP33:%.*]] = shufflevector <128 x i8> [[TMP31]], <128 x i8> [[TMP32]], <128 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15, i32 128, i32 17, i32 18, i32 19, i32 20, i32 21, i32 22, i32 23, i32 24, i32 25, i32 26, i32 27, i32 28, i32 29, i32 30, i32 31, i32 32, i32 33, i32 34, i32 35, i32 36, i32 37, i32 38, i32 39, i32 40, i32 41, i32 42, i32 43, i32 44, i32 45, i32 46, i32 47, i32 48, i32 49, i32 50, i32 51, i32 52, i32 53, i32 54, i32 55, i32 56, i32 57, i32 58, i32 59, i32 60, i32 61, i32 62, i32 63, i32 64, i32 65, i32 66, i32 67, i32 68, i32 69, i32 70, i32 71, i32 72, i32 73, i32 74, i32 75, i32 76, i32 77, i32 78, i32 79, i32 80, i32 81, i32 82, i32 83, i32 84, i32 85, i32 86, i32 87, i32 88, i32 89, i32 90, i32 91, i32 92, i32 93, i32 94, i32 95, i32 96, i32 97, i32 98, i32 99, i32 100, i32 101, i32 102, i32 103, i32 104, i32 105, i32 106, i32 107, i32 108, i32 109, i32 110, i32 111, i32 112, i32 113, i32 114, i32 115, i32 116, i32 117, i32 118, i32 119, i32 120, i32 121, i32 122, i32 123, i32 124, i32 125, i32 126, i32 127>
+; OPT-NEXT: [[TMP34:%.*]] = extractelement <16 x i8> [[IN]], i64 1
+; OPT-NEXT: [[TMP35:%.*]] = insertelement <128 x i8> [[TMP33]], i8 [[TMP34]], i32 17
+; OPT-NEXT: [[TMP36:%.*]] = extractelement <16 x i8> [[IN]], i64 2
+; OPT-NEXT: [[TMP37:%.*]] = insertelement <128 x i8> [[TMP35]], i8 [[TMP36]], i32 18
+; OPT-NEXT: [[TMP38:%.*]] = extractelement <16 x i8> [[IN]], i64 3
+; OPT-NEXT: [[TMP39:%.*]] = insertelement <128 x i8> [[TMP37]], i8 [[TMP38]], i32 19
+; OPT-NEXT: [[TMP40:%.*]] = extractelement <16 x i8> [[IN]], i64 4
+; OPT-NEXT: [[TMP41:%.*]] = insertelement <128 x i8> [[TMP39]], i8 [[TMP40]], i32 20
+; OPT-NEXT: [[TMP42:%.*]] = extractelement <16 x i8> [[IN]], i64 5
+; OPT-NEXT: [[TMP43:%.*]] = insertelement <128 x i8> [[TMP41]], i8 [[TMP42]], i32 21
+; OPT-NEXT: [[TMP44:%.*]] = extractelement <16 x i8> [[IN]], i64 6
+; OPT-NEXT: [[TMP45:%.*]] = insertelement <128 x i8> [[TMP43]], i8 [[TMP44]], i32 22
+; OPT-NEXT: [[TMP46:%.*]] = extractelement <16 x i8> [[IN]], i64 7
+; OPT-NEXT: [[TMP47:%.*]] = insertelement <128 x i8> [[TMP45]], i8 [[TMP46]], i32 23
+; OPT-NEXT: [[TMP48:%.*]] = extractelement <16 x i8> [[IN]], i64 8
+; OPT-NEXT: [[TMP49:%.*]] = insertelement <128 x i8> [[TMP47]], i8 [[TMP48]], i32 24
+; OPT-NEXT: [[TMP50:%.*]] = extractelement <16 x i8> [[IN]], i64 9
+; OPT-NEXT: [[TMP51:%.*]] = insertelement <128 x i8> [[TMP49]], i8 [[TMP50]], i32 25
+; OPT-NEXT: [[TMP52:%.*]] = extractelement <16 x i8> [[IN]], i64 10
+; OPT-NEXT: [[TMP53:%.*]] = insertelement <128 x i8> [[TMP51]], i8 [[TMP52]], i32 26
+; OPT-NEXT: [[TMP54:%.*]] = extractelement <16 x i8> [[IN]], i64 11
+; OPT-NEXT: [[TMP55:%.*]] = insertelement <128 x i8> [[TMP53]], i8 [[TMP54]], i32 27
+; OPT-NEXT: [[TMP56:%.*]] = extractelement <16 x i8> [[IN]], i64 12
+; OPT-NEXT: [[TMP57:%.*]] = insertelement <128 x i8> [[TMP55]], i8 [[TMP56]], i32 28
+; OPT-NEXT: [[TMP58:%.*]] = extractelement <16 x i8> [[IN]], i64 13
+; OPT-NEXT: [[TMP59:%.*]] = insertelement <128 x i8> [[TMP57]], i8 [[TMP58]], i32 29
+; OPT-NEXT: [[TMP60:%.*]] = extractelement <16 x i8> [[IN]], i64 14
+; OPT-NEXT: [[TMP61:%.*]] = insertelement <128 x i8> [[TMP59]], i8 [[TMP60]], i32 30
+; OPT-NEXT: [[TMP62:%.*]] = extractelement <16 x i8> [[IN]], i64 15
+; OPT-NEXT: [[TMP63:%.*]] = insertelement <128 x i8> [[TMP61]], i8 [[TMP62]], i32 31
+; OPT-NEXT: [[TMP64:%.*]] = shufflevector <16 x i8> [[IN]], <16 x i8> poison, <128 x i32> <i32 0, i32 poison, i32 poison, i32 poison, i32 poison, i32 poison, i32 poison, i32 poison, i32 poison, i32 poison, i32 poison, i32 poison, i32 poison, i32 poison, i32 poison, i32 poison, i32 poison, i32 poison, i32 poison, i32 poison, i32 poison, i32 poison, i32 poison, i32 poison, i32 poison, i32 poison, i32 poison, i32 poison, i32 poison, i32 poison, i32 poison, i32 poison, i32 poison, i32 poison, i32 poison, i32 poison, i32 poison, i32 poison, i32 poison, i32 poison, i32 poison, i32 poison, i32 poison, i32 poison, i32 poison, i32 poison, i32 poison, i32 poison, i32 poison, i32 poison, i32 poison, i32 poison, i32 poison, i32 poison, i32 poison, i32 poison, i32 poison, i32 poison, i32 poison, i32 poison, i32 poison, i32 poison, i32 poison, i32 poison, i32 poison, i32 poison, i32 poison, i32 poison, i32 poison, i32 poison, i32 poison, i32 poison, i32 poison, i32 poison, i32 poison, i32 poison, i32 poison, i32 poison, i32 poison, i32 poison, i32 poison, i32 poison, i32 poison, i32 poison, i32 poison, i32 poison, i32 poison, i32 poison, i32 poison, i32 poison, i32 poison, i32 poison, i32 poison, i32 poison, i32 poison, i32 poison, i32 poison, i32 poison, i32 poison, i32 poison, i32 poison, i32 poison, i32 poison, i32 poison, i32 poison, i32 poison, i32 poison, i32 poison, i32 poison, i32 poison, i32 poison, i32 poison, i32 poison, i32 poison, i32 poison, i32 poison, i32 poison, i32 poison, i32 poison, i32 poison, i32 poison, i32 poison, i32 poison, i32 poison, i32 poison, i32 poison, i32 poison, i32 poison>
+; OPT-NEXT: [[TMP65:%.*]] = shufflevector <128 x i8> [[TMP63]], <128 x i8> [[TMP64]], <128 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15, i32 16, i32 17, i32 18, i32 19, i32 20, i32 21, i32 22, i32 23, i32 24, i32 25, i32 26, i32 27, i32 28, i32 29, i32 30, i32 31, i32 128, i32 33, i32 34, i32 35, i32 36, i32 37, i32 38, i32 39, i32 40, i32 41, i32 42, i32 43, i32 44, i32 45, i32 46, i32 47, i32 48, i32 49, i32 50, i32 51, i32 52, i32 53, i32 54, i32 55, i32 56, i32 57, i32 58, i32 59, i32 60, i32 61, i32 62, i32 63, i32 64, i32 65, i32 66, i32 67, i32 68, i32 69, i32 70, i32 71, i32 72, i32 73, i32 74, i32 75, i32 76, i32 77, i32 78, i32 79, i32 80, i32 81, i32 82, i32 83, i32 84, i32 85, i32 86, i32 87, i32 88, i32 89, i32 90, i32 91, i32 92, i32 93, i32 94, i32 95, i32 96, i32 97, i32 98, i32 99, i32 100, i32 101, i32 102, i32 103, i32 104, i32 105, i32 106, i32 107, i32 108, i32 109, i32 110, i32 111, i32 112, i32 113, i32 114, i32 115, i32 116, i32 117, i32 118, i32 119, i32 120, i32 121, i32 122, i32 123, i32 124, i32 125, i32 126, i32 127>
+; OPT-NEXT: [[TMP66:%.*]] = extractelement <16 x i8> [[IN]], i64 1
+; OPT-NEXT: [[TMP67:%.*]] = insertelement <128 x i8> [[TMP65]], i8 [[TMP66]], i32 33
+; OPT-NEXT: [[TMP68:%.*]] = extractelement <16 x i8> [[IN]], i64 2
+; OPT-NEXT: [[TMP69:%.*]] = insertelement <128 x i8> [[TMP67]], i8 [[TMP68]], i32 34
+; OPT-NEXT: [[TMP70:%.*]] = extractelement <16 x i8> [[IN]], i64 3
+; OPT-NEXT: [[TMP71:%.*]] = insertelement <128 x i8> [[TMP69]], i8 [[TMP70]], i32 35
+; OPT-NEXT: [[TMP72:%.*]] = extractelement <16 x i8> [[IN]], i64 4
+; OPT-NEXT: [[TMP73:%.*]] = insertelement <128 x i8> [[TMP71]], i8 [[TMP72]], i32 36
+; OPT-NEXT: [[TMP74:%.*]] = extractelement <16 x i8> [[IN]], i64 5
+; OPT-NEXT: [[TMP75:%.*]] = insertelement <128 x i8> [[TMP73]], i8 [[TMP74]], i32 37
+; OPT-NEXT: [[TMP76:%.*]] = extractelement <16 x i8> [[IN]], i64 6
+; OPT-NEXT: [[TMP77:%.*]] = insertelement <128 x i8> [[TMP75]], i8 [[TMP76]], i32 38
+; OPT-NEXT: [[TMP78:%.*]] = extractelement <16 x i8> [[IN]], i64 7
+; OPT-NEXT: [[TMP79:%.*]] = insertelement <128 x i8> [[TMP77]], i8 [[TMP78]], i32 39
+; OPT-NEXT: [[TMP80:%.*]] = extractelement <16 x i8> [[IN]], i64 8
+; OPT-NEXT: [[TMP81:%.*]] = insertelement <128 x i8> [[TMP79]], i8 [[TMP80]], i32 40
+; OPT-NEXT: [[TMP82:%.*]] = extractelement <16 x i8> [[IN]], i64 9
+; OPT-NEXT: [[TMP83:%.*]] = insertelement <128 x i8> [[TMP81]], i8 [[TMP82]], i32 41
+; OPT-NEXT: [[TMP84:%.*]] = extractelement <16 x i8> [[IN]], i64 10
+; OPT-NEXT: [[TMP85:%.*]] = insertelement <128 x i8> [[TMP83]], i8 [[TMP84]], i32 42
+; OPT-NEXT: [[TMP86:%.*]] = extractelement <16 x i8> [[IN]], i64 11
+; OPT-NEXT: [[TMP87:%.*]] = insertelement <128 x i8> [[TMP85]], i8 [[TMP86]], i32 43
+; OPT-NEXT: [[TMP88:%.*]] = extractelement <16 x i8> [[IN]], i64 12
+; OPT-NEXT: [[TMP89:%.*]] = insertelement <128 x i8> [[TMP87]], i8 [[TMP88]], i32 44
+; OPT-NEXT: [[TMP90:%.*]] = extractelement <16 x i8> [[IN]], i64 13
+; OPT-NEXT: [[TMP91:%.*]] = insertelement <128 x i8> [[TMP89]], i8 [[TMP90]], i32 45
+; OPT-NEXT: [[TMP92:%.*]] = extractelement <16 x i8> [[IN]], i64 14
+; OPT-NEXT: [[TMP93:%.*]] = insertelement <128 x i8> [[TMP91]], i8 [[TMP92]], i32 46
+; OPT-NEXT: [[TMP94:%.*]] = extractelement <16 x i8> [[IN]], i64 15
+; OPT-NEXT: [[TMP95:%.*]] = insertelement <128 x i8> [[TMP93]], i8 [[TMP94]], i32 47
+; OPT-NEXT: [[TMP96:%.*]] = shufflevector <16 x i8> [[IN]], <16 x i8> poison, <128 x i32> <i32 0, i32 poison, i32 poison, i32 poison, i32 poison, i32 poison, i32 poison, i32 poison, i32 poison, i32 poison, i32 poison, i32 poison, i32 poison, i32 poison, i32 poison, i32 poison, i32 poison, i32 poison, i32 poison, i32 poison, i32 poison, i32 poison, i32 poison, i32 poison, i32 poison, i32 poison, i32 poison, i32 poison, i32 poison, i32 poison, i32 poison, i32 poison, i32 poison, i32 poison, i32 poison, i32 poison, i32 poison, i32 poison, i32 poison, i32 poison, i32 poison, i32 poison, i32 poison, i32 poison, i32 poison, i32 poison, i32 poison, i32 poison, i32 poison, i32 poison, i32 poison, i32 poison, i32 poison, i32 poison, i32 poison, i32 poison, i32 poison, i32 poison, i32 poison, i32 poison, i32 poison, i32 poison, i32 poison, i32 poison, i32 poison, i32 poison, i32 poison, i32 poison, i32 poison, i32 poison, i32 poison, i32 poison, i32 poison, i32 poison, i32 poison, i32 poison, i32 poison, i32 poison, i32 poison, i32 poison, i32 poison, i32 poison, i32 poison, i32 poison, i32 poison, i32 poison, i32 poison, i32 poison, i32 poison, i32 poison, i32 poison, i32 poison, i32 poison, i32 poison, i32 poison, i32 poison, i32 poison, i32 poison, i32 poison, i32 poison, i32 poison, i32 poison, i32 poison, i32 poison, i32 poison, i32 poison, i32 poison, i32 poison, i32 poison, i32 poison, i32 poison, i32 poison, i32 poison, i32 poison, i32 poison, i32 poison, i32 poison, i32 poison, i32 poison, i32 poison, i32 poison, i32 poison, i32 poison, i32 poison, i32 poison, i32 poison, i32 poison, i32 poison>
+; OPT-NEXT: [[TMP97:%.*]] = shufflevector <128 x i8> [[TMP95]], <128 x i8> [[TMP96]], <128 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15, i32 16, i32 17, i32 18, i32 19, i32 20, i32 21, i32 22, i32 23, i32 24, i32 25, i32 26, i32 27, i32 28, i32 29, i32 30, i32 31, i32 32, i32 33, i32 34, i32 35, i32 36, i32 37, i32 38, i32 39, i32 40, i32 41, i32 42, i32 43, i32 44, i32 45, i32 46, i32 47, i32 128, i32 49, i32 50, i32 51, i32 52, i32 53, i32 54, i32 55, i32 56, i32 57, i32 58, i32 59, i32 60, i32 61, i32 62, i32 63, i32 64, i32 65, i32 66, i32 67, i32 68, i32 69, i32 70, i32 71, i32 72, i32 73, i32 74, i32 75, i32 76, i32 77, i32 78, i32 79, i32 80, i32 81, i32 82, i32 83, i32 84, i32 85, i32 86, i32 87, i32 88, i32 89, i32 90, i32 91, i32 92, i32 93, i32 94, i32 95, i32 96, i32 97, i32 98, i32 99, i32 100, i32 101, i32 102, i32 103, i32 104, i32 105, i32 106, i32 107, i32 108, i32 109, i32 110, i32 111, i32 112, i32 113, i32 114, i32 115, i32 116, i32 117, i32 118, i32 119, i32 120, i32 121, i32 122, i32 123, i32 124, i32 125, i32 126, i32 127>
+; OPT-NEXT: [[TMP98:%.*]] = extractelement <16 x i8> [[IN]], i64 1
+; OPT-NEXT: [[TMP99:%.*]] = insertelement <128 x i8> [[TMP97]], i8 [[TMP98]], i32 49
+; OPT-NEXT: [[TMP100:%.*]] = extractelement <16 x i8> [[IN]], i64 2
+; OPT-NEXT: [[TMP101:%.*]] = insertelement <128 x i8> [[TMP99]], i8 [[TMP100]], i32 50
+; OPT-NEXT: [[TMP102:%.*]] = extractelement <16 x i8> [[IN]], i64 3
+; OPT-NEXT: [[TMP103:%.*]] = insertelement <128 x i8> [[TMP101]], i8 [[TMP102]], i32 51
+; OPT-NEXT: [[TMP104:%.*]] = extractelement <16 x i8> [[IN]], i64 4
+; OPT-NEXT: [[TMP105:%.*]] = insertelement <128 x i8> [[TMP103]], i8 [[TMP104]], i32 52
+; OPT-NEXT: [[TMP106:%.*]] = extractelement <16 x i8> [[IN]], i64 5
+; OPT-NEXT: [[TMP107:%.*]] = insertelement <128 x i8> [[TMP105]], i8 [[TMP106]], i32 53
+; OPT-NEXT: [[TMP108:%.*]] = extractelement <16 x i8> [[IN]], i64 6
+; OPT-NEXT: [[TMP109:%.*]] = insertelement <128 x i8> [[TMP107]], i8 [[TMP108]], i32 54
+; OPT-NEXT: [[TMP110:%.*]] = extractelement <16 x i8> [[IN]], i64 7
+; OPT-NEXT: [[TMP111:%.*]] = insertelement <128 x i8> [[TMP109]], i8 [[TMP110]], i32 55
+; OPT-NEXT: [[TMP112:%.*]] = extractelement <16 x i8> [[IN]], i64 8
+; OPT-NEXT: [[TMP113:%.*]] = insertelement <128 x i8> [[TMP111]], i8 [[TMP112]], i32 56
+; OPT-NEXT: [[TMP114:%.*]] = extractelement <16 x i8> [[IN]], i64 9
+; OPT-NEXT: [[TMP115:%.*]] = insertelement <128 x i8> [[TMP113]], i8 [[TMP114]], i32 57
+; OPT-NEXT: [[TMP116:%.*]] = extractelement <16 x i8> [[IN]], i64 10
+; OPT-NEXT: [[TMP117:%.*]] = insertelement <128 x i8> [[TMP115]], i8 [[TMP116]], i32 58
+; OPT-NEXT: [[TMP118:%.*]] = extractelement <16 x i8> [[IN]], i64 11
+; OPT-NEXT: [[TMP119:%.*]] = insertelement <128 x i8> [[TMP117]], i8 [[TMP118]], i32 59
+; OPT-NEXT: [[TMP120:%.*]] = extractelement <16 x i8> [[IN]], i64 12
+; OPT-NEXT: [[TMP121:%.*]] = insertelement <128 x i8> [[TMP119]], i8 [[TMP120]], i32 60
+; OPT-NEXT: [[TMP122:%.*]] = extractelement <16 x i8> [[IN]], i64 13
+; OPT-NEXT: [[TMP123:%.*]] = insertelement <128 x i8> [[TMP121]], i8 [[TMP122]], i32 61
+; OPT-NEXT: [[TMP124:%.*]] = extractelement <16 x i8> [[IN]], i64 14
+; OPT-NEXT: [[TMP125:%.*]] = insertelement <128 x i8> [[TMP123]], i8 [[TMP124]], i32 62
+; OPT-NEXT: [[TMP126:%.*]] = extractelement <16 x i8> [[IN]], i64 15
+; OPT-NEXT: [[TMP127:%.*]] = insertelement <128 x i8> [[TMP125]], i8 [[TMP126]], i32 63
+; OPT-NEXT: [[TMP128:%.*]] = shufflevector <16 x i8> [[IN]], <16 x i8> poison, <128 x i32> <i32 0, i32 poison, i32 poison, i32 poison, i32 poison, i32 poison, i32 poison, i32 poison, i32 poison, i32 poison, i32 poison, i32 poison, i32 poison, i32 poison, i32 poison, i32 poison, i32 poison, i32 poison, i32 poison, i32 poison, i32 poison, i32 poison, i32 poison, i32 poison, i32 poison, i32 poison, i32 poison, i32 poison, i32 poison, i32 poison, i32 poison, i32 poison, i32 poison, i32 poison, i32 poison, i32 poison, i32 poison, i32 poison, i32 poison, i32 poison, i32 poison, i32 poison, i32 poison, i32 poison, i32 poison, i32 poison, i32 poison, i32 poison, i32 poison, i32 poison, i32 poison, i32 poison, i32 poison, i32 poison, i32 poison, i32 poison, i32 poison, i32 poison, i32 poison, i32 poison, i32 poison, i32 poison, i32 poison, i32 poison, i32 poison, i32 poison, i32 poison, i32 poison, i32 poison, i32 poison, i32 poison, i32 poison, i32 poison, i32 poison, i32 poison, i32 poison, i32 poison, i32 poison, i32 poison, i32 poison, i32 poison, i32 poison, i32 poison, i32 poison, i32 poison, i32 poison, i32 poison, i32 poison, i32 poison, i32 poison, i32 poison, i32 poison, i32 poison, i32 poison, i32 poison, i32 poison, i32 poison, i32 poison, i32 poison, i32 poison, i32 poison, i32 poison, i32 poison, i32 poison, i32 poison, i32 poison, i32 poison, i32 poison, i32 poison, i32 poison, i32 poison, i32 poison, i32 poison, i32 poison, i32 poison, i32 poison, i32 poison, i32 poison, i32 poison, i32 poison, i32 poison, i32 poison, i32 poison, i32 poison, i32 poison, i32 poison, i32 poison, i32 poison>
+; OPT-NEXT: [[TMP129:%.*]] = shufflevector <128 x i8> [[TMP127]], <128 x i8> [[TMP128]], <128 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15, i32 16, i32 17, i32 18, i32 19, i32 20, i32 21, i32 22, i32 23, i32 24, i32 25, i32 26, i32 27, i32 28, i32 29, i32 30, i32 31, i32 32, i32 33, i32 34, i32 35, i32 36, i32 37, i32 38, i32 39, i32 40, i32 41, i32 42, i32 43, i32 44, i32 45, i32 46, i32 47, i32 48, i32 49, i32 50, i32 51, i32 52, i32 53, i32 54, i32 55, i32 56, i32 57, i32 58, i32 59, i32 60, i32 61, i32 62, i32 63, i32 128, i32 65, i32 66, i32 67, i32 68, i32 69, i32 70, i32 71, i32 72, i32 73, i32 74, i32 75, i32 76, i32 77, i32 78, i32 79, i32 80, i32 81, i32 82, i32 83, i32 84, i32 85, i32 86, i32 87, i32 88, i32 89, i32 90, i32 91, i32 92, i32 93, i32 94, i32 95, i32 96, i32 97, i32 98, i32 99, i32 100, i32 101, i32 102, i32 103, i32 104, i32 105, i32 106, i32 107, i32 108, i32 109, i32 110, i32 111, i32 112, i32 113, i32 114, i32 115, i32 116, i32 117, i32 118, i32 119, i32 120, i32 121, i32 122, i32 123, i32 124, i32 125, i32 126, i32 127>
+; OPT-NEXT: [[TMP130:%.*]] = extractelement <16 x i8> [[IN]], i64 1
+; OPT-NEXT: [[TMP131:%.*]] = insertelement <128 x i8> [[TMP129]], i8 [[TMP130]], i32 65
+; OPT-NEXT: [[TMP132:%.*]] = extractelement <16 x i8> [[IN]], i64 2
+; OPT-NEXT: [[TMP133:%.*]] = insertelement <128 x i8> [[TMP131]], i8 [[TMP132]], i32 66
+; OPT-NEXT: [[TMP134:%.*]] = extractelement <16 x i8> [[IN]], i64 3
+; OPT-NEXT: [[TMP135:%.*]] = insertelement <128 x i8> [[TMP133]], i8 [[TMP134]], i32 67
+; OPT-NEXT: [[TMP136:%.*]] = extractelement <16 x i8> [[IN]], i64 4
+; OPT-NEXT: [[TMP137:%.*]] = insertelement <128 x i8> [[TMP135]], i8 [[TMP136]], i32 68
+; OPT-NEXT: [[TMP138:%.*]] = extractelement <16 x i8> [[IN]], i64 5
+; OPT-NEXT: [[TMP139:%.*]] = insertelement <128 x i8> [[TMP137]], i8 [[TMP138]], i32 69
+; OPT-NEXT: [[TMP140:%.*]] = extractelement <16 x i8> [[IN]], i64 6
+; OPT-NEXT: [[TMP141:%.*]] = insertelement <128 x i8> [[TMP139]], i8 [[TMP140]], i32 70
+; OPT-NEXT: [[TMP142:%.*]] = extractelement <16 x i8> [[IN]], i64 7
+; OPT-NEXT: [[TMP143:%.*]] = insertelement <128 x i8> [[TMP141]], i8 [[TMP142]], i32 71
+; OPT-NEXT: [[TMP144:%.*]] = extractelement <16 x i8> [[IN]], i64 8
+; OPT-NEXT: [[TMP145:%.*]] = insertelement <128 x i8> [[TMP143]], i8 [[TMP144]], i32 72
+; OPT-NEXT: [[TMP146:%.*]] = extractelement <16 x i8> [[IN]], i64 9
+; OPT-NEXT: [[TMP147:%.*]] = insertelement <128 x i8> [[TMP145]], i8 [[TMP146]], i32 73
+; OPT-NEXT: [[TMP148:%.*]] = extractelement <16 x i8> [[IN]], i64 10
+; OPT-NEXT: [[TMP149:%.*]] = insertelement <128 x i8> [[TMP147]], i8 [[TMP148]], i32 74
+; OPT-NEXT: [[TMP150:%.*]] = extractelement <16 x i8> [[IN]], i64 11
+; OPT-NEXT: [[TMP151:%.*]] = insertelement <128 x i8> [[TMP149]], i8 [[TMP150]], i32 75
+; OPT-NEXT: [[TMP152:%.*]] = extractelement <16 x i8> [[IN]], i64 12
+; OPT-NEXT: [[TMP153:%.*]] = insertelement <128 x i8> [[TMP151]], i8 [[TMP152]], i32 76
+; OPT-NEXT: [[TMP154:%.*]] = extractelement <16 x i8> [[IN]], i64 13
+; OPT-NEXT: [[TMP155:%.*]] = insertelement <128 x i8> [[TMP153]], i8 [[TMP154]], i32 77
+; OPT-NEXT: [[TMP156:%.*]] = extractelement <16 x i8> [[IN]], i64 14
+; OPT-NEXT: [[TMP157:%.*]] = insertelement <128 x i8> [[TMP155]], i8 [[TMP156]], i32 78
+; OPT-NEXT: [[TMP158:%.*]] = extractelement <16 x i8> [[IN]], i64 15
+; OPT-NEXT: [[TMP159:%.*]] = insertelement <128 x i8> [[TMP157]], i8 [[TMP158]], i32 79
+; OPT-NEXT: [[TMP160:%.*]] = shufflevector <16 x i8> [[IN]], <16 x i8> poison, <128 x i32> <i32 0, i32 poison, i32 poison, i32 poison, i32 poison, i32 poison, i32 poison, i32 poison, i32 poison, i32 poison, i32 poison, i32 poison, i32 poison, i32 poison, i32 poison, i32 poison, i32 poison, i32 poison, i32 poison, i32 poison, i32 poison, i32 poison, i32 poison, i32 poison, i32 poison, i32 poison, i32 poison, i32 poison, i32 poison, i32 poison, i32 poison, i32 poison, i32 poison, i32 poison, i32 poison, i32 poison, i32 poison, i32 poison, i32 poison, i32 poison, i32 poison, i32 poison, i32 poison, i32 poison, i32 poison, i32 poison, i32 poison, i32 poison, i32 poison, i32 poison, i32 poison, i32 poison, i32 poison, i32 poison, i32 poison, i32 poison, i32 poison, i32 poison, i32 poison, i32 poison, i32 poison, i32 poison, i32 poison, i32 poison, i32 poison, i32 poison, i32 poison, i32 poison, i32 poison, i32 poison, i32 poison, i32 poison, i32 poison, i32 poison, i32 poison, i32 poison, i32 poison, i32 poison, i32 poison, i32 poison, i32 poison, i32 poison, i32 poison, i32 poison, i32 poison, i32 poison, i32 poison, i32 poison, i32 poison, i32 poison, i32 poison, i32 poison, i32 poison, i32 poison, i32 poison, i32 poison, i32 poison, i32 poison, i32 poison, i32 poison, i32 poison, i32 poison, i32 poison, i32 poison, i32 poison, i32 poison, i32 poison, i32 poison, i32 poison, i32 poison, i32 poison, i32 poison, i32 poison, i32 poison, i32 poison, i32 poison, i32 poison, i32 poison, i32 poison, i32 poison, i32 poison, i32 poison, i32 poison, i32 poison, i32 poison, i32 poison, i32 poison, i32 poison>
+; OPT-NEXT: [[TMP161:%.*]] = shufflevector <128 x i8> [[TMP159]], <128 x i8> [[TMP160]], <128 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15, i32 16, i32 17, i32 18, i32 19, i32 20, i32 21, i32 22, i32 23, i32 24, i32 25, i32 26, i32 27, i32 28, i32 29, i32 30, i32 31, i32 32, i32 33, i32 34, i32 35, i32 36, i32 37, i32 38, i32 39, i32 40, i32 41, i32 42, i32 43, i32 44, i32 45, i32 46, i32 47, i32 48, i32 49, i32 50, i32 51, i32 52, i32 53, i32 54, i32 55, i32 56, i32 57, i32 58, i32 59, i32 60, i32 61, i32 62, i32 63, i32 64, i32 65, i32 66, i32 67, i32 68, i32 69, i32 70, i32 71, i32 72, i32 73, i32 74, i32 75, i32 76, i32 77, i32 78, i32 79, i32 128, i32 81, i32 82, i32 83, i32 84, i32 85, i32 86, i32 87, i32 88, i32 89, i32 90, i32 91, i32 92, i32 93, i32 94, i32 95, i32 96, i32 97, i32 98, i32 99, i32 100, i32 101, i32 102, i32 103, i32 104, i32 105, i32 106, i32 107, i32 108, i32 109, i32 110, i32 111, i32 112, i32 113, i32 114, i32 115, i32 116, i32 117, i32 118, i32 119, i32 120, i32 121, i32 122, i32 123, i32 124, i32 125, i32 126, i32 127>
+; OPT-NEXT: [[TMP162:%.*]] = extractelement <16 x i8> [[IN]], i64 1
+; OPT-NEXT: [[TMP163:%.*]] = insertelement <128 x i8> [[TMP161]], i8 [[TMP162]], i32 81
+; OPT-NEXT: [[TMP164:%.*]] = extractelement <16 x i8> [[IN]], i64 2
+; OPT-NEXT: [[TMP165:%.*]] = insertelement <128 x i8> [[TMP163]], i8 [[TMP164]], i32 82
+; OPT-NEXT: [[TMP166:%.*]] = extractelement <16 x i8> [[IN]], i64 3
+; OPT-NEXT: [[TMP167:%.*]] = insertelement <128 x i8> [[TMP165]], i8 [[TMP166]], i32 83
+; OPT-NEXT: [[TMP168:%.*]] = extractelement <16 x i8> [[IN]], i64 4
+; OPT-NEXT: [[TMP169:%.*]] = insertelement <128 x i8> [[TMP167]], i8 [[TMP168]], i32 84
+; OPT-NEXT: [[TMP170:%.*]] = extractelement <16 x i8> [[IN]], i64 5
+; OPT-NEXT: [[TMP171:%.*]] = insertelement <128 x i8> [[TMP169]], i8 [[TMP170]], i32 85
+; OPT-NEXT: [[TMP172:%.*]] = extractelement <16 x i8> [[IN]], i64 6
+; OPT-NEXT: [[TMP173:%.*]] = insertelement <128 x i8> [[TMP171]], i8 [[TMP172]], i32 86
+; OPT-NEXT: [[TMP174:%.*]] = extractelement <16 x i8> [[IN]], i64 7
+; OPT-NEXT: [[TMP175:%.*]] = insertelement <128 x i8> [[TMP173]], i8 [[TMP174]], i32 87
+; OPT-NEXT: [[TMP176:%.*]] = extractelement <16 x i8> [[IN]], i64 8
+; OPT-NEXT: [[TMP177:%.*]] = insertelement <128 x i8> [[TMP175]], i8 [[TMP176]], i32 88
+; OPT-NEXT: [[TMP178:%.*]] = extractelement <16 x i8> [[IN]], i64 9
+; OPT-NEXT: [[TMP179:%.*]] = insertelement <128 x i8> [[TMP177]], i8 [[TMP178]], i32 89
+; OPT-NEXT: [[TMP180:%.*]] = extractelement <16 x i8> [[IN]], i64 10
+; OPT-NEXT: [[TMP181:%.*]] = insertelement <128 x i8> [[TMP179]], i8 [[TMP180]], i32 90
+; OPT-NEXT: [[TMP182:%.*]] = extractelement <16 x i8> [[IN]], i64 11
+; OPT-NEXT: [[TMP183:%.*]] = insertelement <128 x i8> [[TMP181]], i8 [[TMP182]], i32 91
+; OPT-NEXT: [[TMP184:%.*]] = extractelement <16 x i8> [[IN]], i64 12
+; OPT-NEXT: [[TMP185:%.*]] = insertelement <128 x i8> [[TMP183]], i8 [[TMP184]], i32 92
+; OPT-NEXT: [[TMP186:%.*]] = extractelement <16 x i8> [[IN]], i64 13
+; OPT-NEXT: [[TMP187:%.*]] = insertelement <128 x i8> [[TMP185]], i8 [[TMP186]], i32 93
+; OPT-NEXT: [[TMP188:%.*]] = extractelement <16 x i8> [[IN]], i64 14
+; OPT-NEXT: [[TMP189:%.*]] = insertelement <128 x i8> [[TMP187]], i8 [[TMP188]], i32 94
+; OPT-NEXT: [[TMP190:%.*]] = extractelement <16 x i8> [[IN]], i64 15
+; OPT-NEXT: [[TMP191:%.*]] = insertelement <128 x i8> [[TMP189]], i8 [[TMP190]], i32 95
+; OPT-NEXT: [[TMP192:%.*]] = shufflevector <16 x i8> [[IN]], <16 x i8> poison, <128 x i32> <i32 0, i32 poison, i32 poison, i32 poison, i32 poison, i32 poison, i32 poison, i32 poison, i32 poison, i32 poison, i32 poison, i32 poison, i32 poison, i32 poison, i32 poison, i32 poison, i32 poison, i32 poison, i32 poison, i32 poison, i32 poison, i32 poison, i32 poison, i32 poison, i32 poison, i32 poison, i32 poison, i32 poison, i32 poison, i32 poison, i32 poison, i32 poison, i32 poison, i32 poison, i32 poison, i32 poison, i32 poison, i32 poison, i32 poison, i32 poison, i32 poison, i32 poison, i32 poison, i32 poison, i32 poison, i32 poison, i32 poison, i32 poison, i32 poison, i32 poison, i32 poison, i32 poison, i32 poison, i32 poison, i32 poison, i32 poison, i32 poison, i32 poison, i32 poison, i32 poison, i32 poison, i32 poison, i32 poison, i32 poison, i32 poison, i32 poison, i32 poison, i32 poison, i32 poison, i32 poison, i32 poison, i32 poison, i32 poison, i32 poison, i32 poison, i32 poison, i32 poison, i32 poison, i32 poison, i32 poison, i32 poison, i32 poison, i32 poison, i32 poison, i32 poison, i32 poison, i32 poison, i32 poison, i32 poison, i32 poison, i32 poison, i32 poison, i32 poison, i32 poison, i32 poison, i32 poison, i32 poison, i32 poison, i32 poison, i32 poison, i32 poison, i32 poison, i32 poison, i32 poison, i32 poison, i32 poison, i32 poison, i32 poison, i32 poison, i32 poison, i32 poison, i32 poison, i32 poison, i32 poison, i32 poison, i32 poison, i32 poison, i32 poison, i32 poison, i32 poison, i32 poison, i32 poison, i32 poison, i32 poison, i32 poison, i32 poison, i32 poison, i32 poison>
+; OPT-NEXT: [[TMP193:%.*]] = shufflevector <128 x i8> [[TMP191]], <128 x i8> [[TMP192]], <128 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15, i32 16, i32 17, i32 18, i32 19, i32 20, i32 21, i32 22, i32 23, i32 24, i32 25, i32 26, i32 27, i32 28, i32 29, i32 30, i32 31, i32 32, i32 33, i32 34, i32 35, i32 36, i32 37, i32 38, i32 39, i32 40, i32 41, i32 42, i32 43, i32 44, i32 45, i32 46, i32 47, i32 48, i32 49, i32 50, i32 51, i32 52, i32 53, i32 54, i32 55, i32 56, i32 57, i32 58, i32 59, i32 60, i32 61, i32 62, i32 63, i32 64, i32 65, i32 66, i32 67, i32 68, i32 69, i32 70, i32 71, i32 72, i32 73, i32 74, i32 75, i32 76, i32 77, i32 78, i32 79, i32 80, i32 81, i32 82, i32 83, i32 84, i32 85, i32 86, i32 87, i32 88, i32 89, i32 90, i32 91, i32 92, i32 93, i32 94, i32 95, i32 128, i32 97, i32 98, i32 99, i32 100, i32 101, i32 102, i32 103, i32 104, i32 105, i32 106, i32 107, i32 108, i32 109, i32 110, i32 111, i32 112, i32 113, i32 114, i32 115, i32 116, i32 117, i32 118, i32 119, i32 120, i32 121, i32 122, i32 123, i32 124, i32 125, i32 126, i32 127>
+; OPT-NEXT: [[TMP194:%.*]] = extractelement <16 x i8> [[IN]], i64 1
+; OPT-NEXT: [[TMP195:%.*]] = insertelement <128 x i8> [[TMP193]], i8 [[TMP194]], i32 97
+; OPT-NEXT: [[TMP196:%.*]] = extractelement <16 x i8> [[IN]], i64 2
+; OPT-NEXT: [[TMP197:%.*]] = insertelement <128 x i8> [[TMP195]], i8 [[TMP196]], i32 98
+; OPT-NEXT: [[TMP198:%.*]] = extractelement <16 x i8> [[IN]], i64 3
+; OPT-NEXT: [[TMP199:%.*]] = insertelement <128 x i8> [[TMP197]], i8 [[TMP198]], i32 99
+; OPT-NEXT: [[TMP200:%.*]] = extractelement <16 x i8> [[IN]], i64 4
+; OPT-NEXT: [[TMP201:%.*]] = insertelement <128 x i8> [[TMP199]], i8 [[TMP200]], i32 100
+; OPT-NEXT: [[TMP202:%.*]] = extractelement <16 x i8> [[IN]], i64 5
+; OPT-NEXT: [[TMP203:%.*]] = insertelement <128 x i8> [[TMP201]], i8 [[TMP202]], i32 101
+; OPT-NEXT: [[TMP204:%.*]] = extractelement <16 x i8> [[IN]], i64 6
+; OPT-NEXT: [[TMP205:%.*]] = insertelement <128 x i8> [[TMP203]], i8 [[TMP204]], i32 102
+; OPT-NEXT: [[TMP206:%.*]] = extractelement <16 x i8> [[IN]], i64 7
+; OPT-NEXT: [[TMP207:%.*]] = insertelement <128 x i8> [[TMP205]], i8 [[TMP206]], i32 103
+; OPT-NEXT: [[TMP208:%.*]] = extractelement <16 x i8> [[IN]], i64 8
+; OPT-NEXT: [[TMP209:%.*]] = insertelement <128 x i8> [[TMP207]], i8 [[TMP208]], i32 104
+; OPT-NEXT: [[TMP210:%.*]] = extractelement <16 x i8> [[IN]], i64 9
+; OPT-NEXT: [[TMP211:%.*]] = insertelement <128 x i8> [[TMP209]], i8 [[TMP210]], i32 105
+; OPT-NEXT: [[TMP212:%.*]] = extractelement <16 x i8> [[IN]], i64 10
+; OPT-NEXT: [[TMP213:%.*]] = insertelement <128 x i8> [[TMP211]], i8 [[TMP212]], i32 106
+; OPT-NEXT: [[TMP214:%.*]] = extractelement <16 x i8> [[IN]], i64 11
+; OPT-NEXT: [[TMP215:%.*]] = insertelement <128 x i8> [[TMP213]], i8 [[TMP214]], i32 107
+; OPT-NEXT: [[TMP216:%.*]] = extractelement <16 x i8> [[IN]], i64 12
+; OPT-NEXT: [[TMP217:%.*]] = insertelement <128 x i8> [[TMP215]], i8 [[TMP216]], i32 108
+; OPT-NEXT: [[TMP218:%.*]] = extractelement <16 x i8> [[IN]], i64 13
+; OPT-NEXT: [[TMP219:%.*]] = insertelement <128 x i8> [[TMP217]], i8 [[TMP218]], i32 109
+; OPT-NEXT: [[TMP220:%.*]] = extractelement <16 x i8> [[IN]], i64 14
+; OPT-NEXT: [[TMP221:%.*]] = insertelement <128 x i8> [[TMP219]], i8 [[TMP220]], i32 110
+; OPT-NEXT: [[TMP222:%.*]] = extractelement <16 x i8> [[IN]], i64 15
+; OPT-NEXT: [[TMP223:%.*]] = insertelement <128 x i8> [[TMP221]], i8 [[TMP222]], i32 111
+; OPT-NEXT: [[TMP224:%.*]] = shufflevector <16 x i8> [[IN]], <16 x i8> poison, <128 x i32> <i32 0, i32 poison, i32 poison, i32 poison, i32 poison, i32 poison, i32 poison, i32 poison, i32 poison, i32 poison, i32 poison, i32 poison, i32 poison, i32 poison, i32 poison, i32 poison, i32 poison, i32 poison, i32 poison, i32 poison, i32 poison, i32 poison, i32 poison, i32 poison, i32 poison, i32 poison, i32 poison, i32 poison, i32 poison, i32 poison, i32 poison, i32 poison, i32 poison, i32 poison, i32 poison, i32 poison, i32 poison, i32 poison, i32 poison, i32 poison, i32 poison, i32 poison, i32 poison, i32 poison, i32 poison, i32 poison, i32 poison, i32 poison, i32 poison, i32 poison, i32 poison, i32 poison, i32 poison, i32 poison, i32 poison, i32 poison, i32 poison, i32 poison, i32 poison, i32 poison, i32 poison, i32 poison, i32 poison, i32 poison, i32 poison, i32 poison, i32 poison, i32 poison, i32 poison, i32 poison, i32 poison, i32 poison, i32 poison, i32 poison, i32 poison, i32 poison, i32 poison, i32 poison, i32 poison, i32 poison, i32 poison, i32 poison, i32 poison, i32 poison, i32 poison, i32 poison, i32 poison, i32 poison, i32 poison, i32 poison, i32 poison, i32 poison, i32 poison, i32 poison, i32 poison, i32 poison, i32 poison, i32 poison, i32 poison, i32 poison, i32 poison, i32 poison, i32 poison, i32 poison, i32 poison, i32 poison, i32 poison, i32 poison, i32 poison, i32 poison, i32 poison, i32 poison, i32 poison, i32 poison, i32 poison, i32 poison, i32 poison, i32 poison, i32 poison, i32 poison, i32 poison, i32 poison, i32 poison, i32 poison, i32 poison, i32 poison, i32 poison, i32 poison>
+; OPT-NEXT: [[TMP225:%.*]] = shufflevector <128 x i8> [[TMP223]], <128 x i8> [[TMP224]], <128 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15, i32 16, i32 17, i32 18, i32 19, i32 20, i32 21, i32 22, i32 23, i32 24, i32 25, i32 26, i32 27, i32 28, i32 29, i32 30, i32 31, i32 32, i32 33, i32 34, i32 35, i32 36, i32 37, i32 38, i32 39, i32 40, i32 41, i32 42, i32 43, i32 44, i32 45, i32 46, i32 47, i32 48, i32 49, i32 50, i32 51, i32 52, i32 53, i32 54, i32 55, i32 56, i32 57, i32 58, i32 59, i32 60, i32 61, i32 62, i32 63, i32 64, i32 65, i32 66, i32 67, i32 68, i32 69, i32 70, i32 71, i32 72, i32 73, i32 74, i32 75, i32 76, i32 77, i32 78, i32 79, i32 80, i32 81, i32 82, i32 83, i32 84, i32 85, i32 86, i32 87, i32 88, i32 89, i32 90, i32 91, i32 92, i32 93, i32 94, i32 95, i32 96, i32 97, i32 98, i32 99, i32 100, i32 101, i32 102, i32 103, i32 104, i32 105, i32 106, i32 107, i32 108, i32 109, i32 110, i32 111, i32 128, i32 113, i32 114, i32 115, i32 116, i32 117, i32 118, i32 119, i32 120, i32 121, i32 122, i32 123, i32 124, i32 125, i32 126, i32 127>
+; OPT-NEXT: [[TMP226:%.*]] = extractelement <16 x i8> [[IN]], i64 1
+; OPT-NEXT: [[TMP227:%.*]] = insertelement <128 x i8> [[TMP225]], i8 [[TMP226]], i32 113
+; OPT-NEXT: [[TMP228:%.*]] = extractelement <16 x i8> [[IN]], i64 2
+; OPT-NEXT: [[TMP229:%.*]] = insertelement <128 x i8> [[TMP227]], i8 [[TMP228]], i32 114
+; OPT-NEXT: [[TMP230:%.*]] = extractelement <16 x i8> [[IN]], i64 3
+; OPT-NEXT: [[TMP231:%.*]] = insertelement <128 x i8> [[TMP229]], i8 [[TMP230]], i32 115
+; OPT-NEXT: [[TMP232:%.*]] = extractelement <16 x i8> [[IN]], i64 4
+; OPT-NEXT: [[TMP233:%.*]] = insertelement <128 x i8> [[TMP231]], i8 [[TMP232]], i32 116
+; OPT-NEXT: [[TMP234:%.*]] = extractelement <16 x i8> [[IN]], i64 5
+; OPT-NEXT: [[TMP235:%.*]] = insertelement <128 x i8> [[TMP233]], i8 [[TMP234]], i32 117
+; OPT-NEXT: [[TMP236:%.*]] = extractelement <16 x i8> [[IN]], i64 6
+; OPT-NEXT: [[TMP237:%.*]] = insertelement <128 x i8> [[TMP235]], i8 [[TMP236]], i32 118
+; OPT-NEXT: [[TMP238:%.*]] = extractelement <16 x i8> [[IN]], i64 7
+; OPT-NEXT: [[TMP239:%.*]] = insertelement <128 x i8> [[TMP237]], i8 [[TMP238]], i32 119
+; OPT-NEXT: [[TMP240:%.*]] = extractelement <16 x i8> [[IN]], i64 8
+; OPT-NEXT: [[TMP241:%.*]] = insertelement <128 x i8> [[TMP239]], i8 [[TMP240]], i32 120
+; OPT-NEXT: [[TMP242:%.*]] = extractelement <16 x i8> [[IN]], i64 9
+; OPT-NEXT: [[TMP243:%.*]] = insertelement <128 x i8> [[TMP241]], i8 [[TMP242]], i32 121
+; OPT-NEXT: [[TMP244:%.*]] = extractelement <16 x i8> [[IN]], i64 10
+; OPT-NEXT: [[TMP245:%.*]] = insertelement <128 x i8> [[TMP243]], i8 [[TMP244]], i32 122
+; OPT-NEXT: [[TMP246:%.*]] = extractelement <16 x i8> [[IN]], i64 11
+; OPT-NEXT: [[TMP247:%.*]] = insertelement <128 x i8> [[TMP245]], i8 [[TMP246]], i32 123
+; OPT-NEXT: [[TMP248:%.*]] = extractelement <16 x i8> [[IN]], i64 12
+; OPT-NEXT: [[TMP249:%.*]] = insertelement <128 x i8> [[TMP247]], i8 [[TMP248]], i32 124
+; OPT-NEXT: [[TMP250:%.*]] = extractelement <16 x i8> [[IN]], i64 13
+; OPT-NEXT: [[TMP251:%.*]] = insertelement <128 x i8> [[TMP249]], i8 [[TMP250]], i32 125
+; OPT-NEXT: [[TMP252:%.*]] = extractelement <16 x i8> [[IN]], i64 14
+; OPT-NEXT: [[TMP253:%.*]] = insertelement <128 x i8> [[TMP251]], i8 [[TMP252]], i32 126
+; OPT-NEXT: [[TMP254:%.*]] = extractelement <16 x i8> [[IN]], i64 15
+; OPT-NEXT: [[TMP255:%.*]] = insertelement <128 x i8> [[TMP253]], i8 [[TMP254]], i32 127
+; OPT-NEXT: [[TMP256:%.*]] = shufflevector <16 x i8> [[IN]], <16 x i8> poison, <16 x i32> <i32 0, i32 poison, i32 poison, i32 poison, i32 poison, i32 poison, i32 poison, i32 poison, i32 poison, i32 poison, i32 poison, i32 poison, i32 poison, i32 poison, i32 poison, i32 poison>
+; OPT-NEXT: [[TMP257:%.*]] = insertelement <16 x i8> [[TMP256]], i8 [[TMP162]], i64 1
+; OPT-NEXT: [[TMP258:%.*]] = insertelement <16 x i8> [[TMP257]], i8 [[TMP164]], i64 2
+; OPT-NEXT: [[TMP259:%.*]] = insertelement <16 x i8> [[TMP258]], i8 [[TMP166]], i64 3
+; OPT-NEXT: [[TMP260:%.*]] = insertelement <16 x i8> [[TMP259]], i8 [[TMP168]], i64 4
+; OPT-NEXT: [[TMP261:%.*]] = insertelement <16 x i8> [[TMP260]], i8 [[TMP170]], i64 5
+; OPT-NEXT: [[TMP262:%.*]] = insertelement <16 x i8> [[TMP261]], i8 [[TMP172]], i64 6
+; OPT-NEXT: [[TMP263:%.*]] = insertelement <16 x i8> [[TMP262]], i8 [[TMP174]], i64 7
+; OPT-NEXT: [[TMP264:%.*]] = insertelement <16 x i8> [[TMP263]], i8 [[TMP176]], i64 8
+; OPT-NEXT: [[TMP265:%.*]] = insertelement <16 x i8> [[TMP264]], i8 [[TMP178]], i64 9
+; OPT-NEXT: [[TMP266:%.*]] = insertelement <16 x i8> [[TMP265]], i8 [[TMP180]], i64 10
+; OPT-NEXT: [[TMP267:%.*]] = insertelement <16 x i8> [[TMP266]], i8 [[TMP182]], i64 11
+; OPT-NEXT: [[TMP268:%.*]] = insertelement <16 x i8> [[TMP267]], i8 [[TMP184]], i64 12
+; OPT-NEXT: [[TMP269:%.*]] = insertelement <16 x i8> [[TMP268]], i8 [[TMP186]], i64 13
+; OPT-NEXT: [[TMP270:%.*]] = insertelement <16 x i8> [[TMP269]], i8 [[TMP188]], i64 14
+; OPT-NEXT: [[TMP271:%.*]] = shufflevector <16 x i8> [[TMP270]], <16 x i8> [[IN]], <16 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 31>
+; OPT-NEXT: [[SUM:%.*]] = add <16 x i8> [[TMP271]], [[ADD]]
+; OPT-NEXT: store <16 x i8> [[SUM]], ptr addrspace(3) [[OUT]], align 16
+; OPT-NEXT: ret void
+;
+entry:
+ %alloca = freeze <128 x i8> poison
+ %0 = extractelement <16 x i8> %in, i64 0
+ %1 = insertelement <128 x i8> %alloca, i8 %0, i32 0
+ %2 = extractelement <16 x i8> %in, i64 1
+ %3 = insertelement <128 x i8> %1, i8 %2, i32 1
+ %4 = extractelement <16 x i8> %in, i64 2
+ %5 = insertelement <128 x i8> %3, i8 %4, i32 2
+ %6 = extractelement <16 x i8> %in, i64 3
+ %7 = insertelement <128 x i8> %5, i8 %6, i32 3
+ %8 = extractelement <16 x i8> %in, i64 4
+ %9 = insertelement <128 x i8> %7, i8 %8, i32 4
+ %10 = extractelement <16 x i8> %in, i64 5
+ %11 = insertelement <128 x i8> %9, i8 %10, i32 5
+ %12 = extractelement <16 x i8> %in, i64 6
+ %13 = insertelement <128 x i8> %11, i8 %12, i32 6
+ %14 = extractelement <16 x i8> %in, i64 7
+ %15 = insertelement <128 x i8> %13, i8 %14, i32 7
+ %16 = extractelement <16 x i8> %in, i64 8
+ %17 = insertelement <128 x i8> %15, i8 %16, i32 8
+ %18 = extractelement <16 x i8> %in, i64 9
+ %19 = insertelement <128 x i8> %17, i8 %18, i32 9
+ %20 = extractelement <16 x i8> %in, i64 10
+ %21 = insertelement <128 x i8> %19, i8 %20, i32 10
+ %22 = extractelement <16 x i8> %in, i64 11
+ %23 = insertelement <128 x i8> %21, i8 %22, i32 11
+ %24 = extractelement <16 x i8> %in, i64 12
+ %25 = insertelement <128 x i8> %23, i8 %24, i32 12
+ %26 = extractelement <16 x i8> %in, i64 13
+ %27 = insertelement <128 x i8> %25, i8 %26, i32 13
+ %28 = extractelement <16 x i8> %in, i64 14
+ %29 = insertelement <128 x i8> %27, i8 %28, i32 14
+ %30 = extractelement <16 x i8> %in, i64 15
+ %31 = insertelement <128 x i8> %29, i8 %30, i32 15
+ %32 = extractelement <16 x i8> %in, i64 0
+ %33 = insertelement <128 x i8> %31, i8 %32, i32 16
+ %34 = extractelement <16 x i8> %in, i64 1
+ %35 = insertelement <128 x i8> %33, i8 %34, i32 17
+ %36 = extractelement <16 x i8> %in, i64 2
+ %37 = insertelement <128 x i8> %35, i8 %36, i32 18
+ %38 = extractelement <16 x i8> %in, i64 3
+ %39 = insertelement <128 x i8> %37, i8 %38, i32 19
+ %40 = extractelement <16 x i8> %in, i64 4
+ %41 = insertelement <128 x i8> %39, i8 %40, i32 20
+ %42 = extractelement <16 x i8> %in, i64 5
+ %43 = insertelement <128 x i8> %41, i8 %42, i32 21
+ %44 = extractelement <16 x i8> %in, i64 6
+ %45 = insertelement <128 x i8> %43, i8 %44, i32 22
+ %46 = extractelement <16 x i8> %in, i64 7
+ %47 = insertelement <128 x i8> %45, i8 %46, i32 23
+ %48 = extractelement <16 x i8> %in, i64 8
+ %49 = insertelement <128 x i8> %47, i8 %48, i32 24
+ %50 = extractelement <16 x i8> %in, i64 9
+ %51 = insertelement <128 x i8> %49, i8 %50, i32 25
+ %52 = extractelement <16 x i8> %in, i64 10
+ %53 = insertelement <128 x i8> %51, i8 %52, i32 26
+ %54 = extractelement <16 x i8> %in, i64 11
+ %55 = insertelement <128 x i8> %53, i8 %54, i32 27
+ %56 = extractelement <16 x i8> %in, i64 12
+ %57 = insertelement <128 x i8> %55, i8 %56, i32 28
+ %58 = extractelement <16 x i8> %in, i64 13
+ %59 = insertelement <128 x i8> %57, i8 %58, i32 29
+ %60 = extractelement <16 x i8> %in, i64 14
+ %61 = insertelement <128 x i8> %59, i8 %60, i32 30
+ %62 = extractelement <16 x i8> %in, i64 15
+ %63 = insertelement <128 x i8> %61, i8 %62, i32 31
+ %64 = extractelement <16 x i8> %in, i64 0
+ %65 = insertelement <128 x i8> %63, i8 %64, i32 32
+ %66 = extractelement <16 x i8> %in, i64 1
+ %67 = insertelement <128 x i8> %65, i8 %66, i32 33
+ %68 = extractelement <16 x i8> %in, i64 2
+ %69 = insertelement <128 x i8> %67, i8 %68, i32 34
+ %70 = extractelement <16 x i8> %in, i64 3
+ %71 = insertelement <128 x i8> %69, i8 %70, i32 35
+ %72 = extractelement <16 x i8> %in, i64 4
+ %73 = insertelement <128 x i8> %71, i8 %72, i32 36
+ %74 = extractelement <16 x i8> %in, i64 5
+ %75 = insertelement <128 x i8> %73, i8 %74, i32 37
+ %76 = extractelement <16 x i8> %in, i64 6
+ %77 = insertelement <128 x i8> %75, i8 %76, i32 38
+ %78 = extractelement <16 x i8> %in, i64 7
+ %79 = insertelement <128 x i8> %77, i8 %78, i32 39
+ %80 = extractelement <16 x i8> %in, i64 8
+ %81 = insertelement <128 x i8> %79, i8 %80, i32 40
+ %82 = extractelement <16 x i8> %in, i64 9
+ %83 = insertelement <128 x i8> %81, i8 %82, i32 41
+ %84 = extractelement <16 x i8> %in, i64 10
+ %85 = insertelement <128 x i8> %83, i8 %84, i32 42
+ %86 = extractelement <16 x i8> %in, i64 11
+ %87 = insertelement <128 x i8> %85, i8 %86, i32 43
+ %88 = extractelement <16 x i8> %in, i64 12
+ %89 = insertelement <128 x i8> %87, i8 %88, i32 44
+ %90 = extractelement <16 x i8> %in, i64 13
+ %91 = insertelement <128 x i8> %89, i8 %90, i32 45
+ %92 = extractelement <16 x i8> %in, i64 14
+ %93 = insertelement <128 x i8> %91, i8 %92, i32 46
+ %94 = extractelement <16 x i8> %in, i64 15
+ %95 = insertelement <128 x i8> %93, i8 %94, i32 47
+ %96 = extractelement <16 x i8> %in, i64 0
+ %97 = insertelement <128 x i8> %95, i8 %96, i32 48
+ %98 = extractelement <16 x i8> %in, i64 1
+ %99 = insertelement <128 x i8> %97, i8 %98, i32 49
+ %100 = extractelement <16 x i8> %in, i64 2
+ %101 = insertelement <128 x i8> %99, i8 %100, i32 50
+ %102 = extractelement <16 x i8> %in, i64 3
+ %103 = insertelement <128 x i8> %101, i8 %102, i32 51
+ %104 = extractelement <16 x i8> %in, i64 4
+ %105 = insertelement <128 x i8> %103, i8 %104, i32 52
+ %106 = extractelement <16 x i8> %in, i64 5
+ %107 = insertelement <128 x i8> %105, i8 %106, i32 53
+ %108 = extractelement <16 x i8> %in, i64 6
+ %109 = insertelement <128 x i8> %107, i8 %108, i32 54
+ %110 = extractelement <16 x i8> %in, i64 7
+ %111 = insertelement <128 x i8> %109, i8 %110, i32 55
+ %112 = extractelement <16 x i8> %in, i64 8
+ %113 = insertelement <128 x i8> %111, i8 %112, i32 56
+ %114 = extractelement <16 x i8> %in, i64 9
+ %115 = insertelement <128 x i8> %113, i8 %114, i32 57
+ %116 = extractelement <16 x i8> %in, i64 10
+ %117 = insertelement <128 x i8> %115, i8 %116, i32 58
+ %118 = extractelement <16 x i8> %in, i64 11
+ %119 = insertelement <128 x i8> %117, i8 %118, i32 59
+ %120 = extractelement <16 x i8> %in, i64 12
+ %121 = insertelement <128 x i8> %119, i8 %120, i32 60
+ %122 = extractelement <16 x i8> %in, i64 13
+ %123 = insertelement <128 x i8> %121, i8 %122, i32 61
+ %124 = extractelement <16 x i8> %in, i64 14
+ %125 = insertelement <128 x i8> %123, i8 %124, i32 62
+ %126 = extractelement <16 x i8> %in, i64 15
+ %127 = insertelement <128 x i8> %125, i8 %126, i32 63
+ %128 = extractelement <16 x i8> %in, i64 0
+ %129 = insertelement <128 x i8> %127, i8 %128, i32 64
+ %130 = extractelement <16 x i8> %in, i64 1
+ %131 = insertelement <128 x i8> %129, i8 %130, i32 65
+ %132 = extractelement <16 x i8> %in, i64 2
+ %133 = insertelement <128 x i8> %131, i8 %132, i32 66
+ %134 = extractelement <16 x i8> %in, i64 3
+ %135 = insertelement <128 x i8> %133, i8 %134, i32 67
+ %136 = extractelement <16 x i8> %in, i64 4
+ %137 = insertelement <128 x i8> %135, i8 %136, i32 68
+ %138 = extractelement <16 x i8> %in, i64 5
+ %139 = insertelement <128 x i8> %137, i8 %138, i32 69
+ %140 = extractelement <16 x i8> %in, i64 6
+ %141 = insertelement <128 x i8> %139, i8 %140, i32 70
+ %142 = extractelement <16 x i8> %in, i64 7
+ %143 = insertelement <128 x i8> %141, i8 %142, i32 71
+ %144 = extractelement <16 x i8> %in, i64 8
+ %145 = insertelement <128 x i8> %143, i8 %144, i32 72
+ %146 = extractelement <16 x i8> %in, i64 9
+ %147 = insertelement <128 x i8> %145, i8 %146, i32 73
+ %148 = extractelement <16 x i8> %in, i64 10
+ %149 = insertelement <128 x i8> %147, i8 %148, i32 74
+ %150 = extractelement <16 x i8> %in, i64 11
+ %151 = insertelement <128 x i8> %149, i8 %150, i32 75
+ %152 = extractelement <16 x i8> %in, i64 12
+ %153 = insertelement <128 x i8> %151, i8 %152, i32 76
+ %154 = extractelement <16 x i8> %in, i64 13
+ %155 = insertelement <128 x i8> %153, i8 %154, i32 77
+ %156 = extractelement <16 x i8> %in, i64 14
+ %157 = insertelement <128 x i8> %155, i8 %156, i32 78
+ %158 = extractelement <16 x i8> %in, i64 15
+ %159 = insertelement <128 x i8> %157, i8 %158, i32 79
+ %160 = extractelement <16 x i8> %in, i64 0
+ %161 = insertelement <128 x i8> %159, i8 %160, i32 80
+ %162 = extractelement <16 x i8> %in, i64 1
+ %163 = insertelement <128 x i8> %161, i8 %162, i32 81
+ %164 = extractelement <16 x i8> %in, i64 2
+ %165 = insertelement <128 x i8> %163, i8 %164, i32 82
+ %166 = extractelement <16 x i8> %in, i64 3
+ %167 = insertelement <128 x i8> %165, i8 %166, i32 83
+ %168 = extractelement <16 x i8> %in, i64 4
+ %169 = insertelement <128 x i8> %167, i8 %168, i32 84
+ %170 = extractelement <16 x i8> %in, i64 5
+ %171 = insertelement <128 x i8> %169, i8 %170, i32 85
+ %172 = extractelement <16 x i8> %in, i64 6
+ %173 = insertelement <128 x i8> %171, i8 %172, i32 86
+ %174 = extractelement <16 x i8> %in, i64 7
+ %175 = insertelement <128 x i8> %173, i8 %174, i32 87
+ %176 = extractelement <16 x i8> %in, i64 8
+ %177 = insertelement <128 x i8> %175, i8 %176, i32 88
+ %178 = extractelement <16 x i8> %in, i64 9
+ %179 = insertelement <128 x i8> %177, i8 %178, i32 89
+ %180 = extractelement <16 x i8> %in, i64 10
+ %181 = insertelement <128 x i8> %179, i8 %180, i32 90
+ %182 = extractelement <16 x i8> %in, i64 11
+ %183 = insertelement <128 x i8> %181, i8 %182, i32 91
+ %184 = extractelement <16 x i8> %in, i64 12
+ %185 = insertelement <128 x i8> %183, i8 %184, i32 92
+ %186 = extractelement <16 x i8> %in, i64 13
+ %187 = insertelement <128 x i8> %185, i8 %186, i32 93
+ %188 = extractelement <16 x i8> %in, i64 14
+ %189 = insertelement <128 x i8> %187, i8 %188, i32 94
+ %190 = extractelement <16 x i8> %in, i64 15
+ %191 = insertelement <128 x i8> %189, i8 %190, i32 95
+ %192 = extractelement <16 x i8> %in, i64 0
+ %193 = insertelement <128 x i8> %191, i8 %192, i32 96
+ %194 = extractelement <16 x i8> %in, i64 1
+ %195 = insertelement <128 x i8> %193, i8 %194, i32 97
+ %196 = extractelement <16 x i8> %in, i64 2
+ %197 = insertelement <128 x i8> %195, i8 %196, i32 98
+ %198 = extractelement <16 x i8> %in, i64 3
+ %199 = insertelement <128 x i8> %197, i8 %198, i32 99
+ %200 = extractelement <16 x i8> %in, i64 4
+ %201 = insertelement <128 x i8> %199, i8 %200, i32 100
+ %202 = extractelement <16 x i8> %in, i64 5
+ %203 = insertelement <128 x i8> %201, i8 %202, i32 101
+ %204 = extractelement <16 x i8> %in, i64 6
+ %205 = insertelement <128 x i8> %203, i8 %204, i32 102
+ %206 = extractelement <16 x i8> %in, i64 7
+ %207 = insertelement <128 x i8> %205, i8 %206, i32 103
+ %208 = extractelement <16 x i8> %in, i64 8
+ %209 = insertelement <128 x i8> %207, i8 %208, i32 104
+ %210 = extractelement <16 x i8> %in, i64 9
+ %211 = insertelement <128 x i8> %209, i8 %210, i32 105
+ %212 = extractelement <16 x i8> %in, i64 10
+ %213 = insertelement <128 x i8> %211, i8 %212, i32 106
+ %214 = extractelement <16 x i8> %in, i64 11
+ %215 = insertelement <128 x i8> %213, i8 %214, i32 107
+ %216 = extractelement <16 x i8> %in, i64 12
+ %217 = insertelement <128 x i8> %215, i8 %216, i32 108
+ %218 = extractelement <16 x i8> %in, i64 13
+ %219 = insertelement <128 x i8> %217, i8 %218, i32 109
+ %220 = extractelement <16 x i8> %in, i64 14
+ %221 = insertelement <128 x i8> %219, i8 %220, i32 110
+ %222 = extractelement <16 x i8> %in, i64 15
+ %223 = insertelement <128 x i8> %221, i8 %222, i32 111
+ %224 = extractelement <16 x i8> %in, i64 0
+ %225 = insertelement <128 x i8> %223, i8 %224, i32 112
+ %226 = extractelement <16 x i8> %in, i64 1
+ %227 = insertelement <128 x i8> %225, i8 %226, i32 113
+ %228 = extractelement <16 x i8> %in, i64 2
+ %229 = insertelement <128 x i8> %227, i8 %228, i32 114
+ %230 = extractelement <16 x i8> %in, i64 3
+ %231 = insertelement <128 x i8> %229, i8 %230, i32 115
+ %232 = extractelement <16 x i8> %in, i64 4
+ %233 = insertelement <128 x i8> %231, i8 %232, i32 116
+ %234 = extractelement <16 x i8> %in, i64 5
+ %235 = insertelement <128 x i8> %233, i8 %234, i32 117
+ %236 = extractelement <16 x i8> %in, i64 6
+ %237 = insertelement <128 x i8> %235, i8 %236, i32 118
+ %238 = extractelement <16 x i8> %in, i64 7
+ %239 = insertelement <128 x i8> %237, i8 %238, i32 119
+ %240 = extractelement <16 x i8> %in, i64 8
+ %241 = insertelement <128 x i8> %239, i8 %240, i32 120
+ %242 = extractelement <16 x i8> %in, i64 9
+ %243 = insertelement <128 x i8> %241, i8 %242, i32 121
+ %244 = extractelement <16 x i8> %in, i64 10
+ %245 = insertelement <128 x i8> %243, i8 %244, i32 122
+ %246 = extractelement <16 x i8> %in, i64 11
+ %247 = insertelement <128 x i8> %245, i8 %246, i32 123
+ %248 = extractelement <16 x i8> %in, i64 12
+ %249 = insertelement <128 x i8> %247, i8 %248, i32 124
+ %250 = extractelement <16 x i8> %in, i64 13
+ %251 = insertelement <128 x i8> %249, i8 %250, i32 125
+ %252 = extractelement <16 x i8> %in, i64 14
+ %253 = insertelement <128 x i8> %251, i8 %252, i32 126
+ %254 = extractelement <16 x i8> %in, i64 15
+ %255 = insertelement <128 x i8> %253, i8 %254, i32 127
+ %256 = insertelement <16 x i8> poison, i8 %160, i64 0
+ %257 = insertelement <16 x i8> %256, i8 %162, i64 1
+ %258 = insertelement <16 x i8> %257, i8 %164, i64 2
+ %259 = insertelement <16 x i8> %258, i8 %166, i64 3
+ %260 = insertelement <16 x i8> %259, i8 %168, i64 4
+ %261 = insertelement <16 x i8> %260, i8 %170, i64 5
+ %262 = insertelement <16 x i8> %261, i8 %172, i64 6
+ %263 = insertelement <16 x i8> %262, i8 %174, i64 7
+ %264 = insertelement <16 x i8> %263, i8 %176, i64 8
+ %265 = insertelement <16 x i8> %264, i8 %178, i64 9
+ %266 = insertelement <16 x i8> %265, i8 %180, i64 10
+ %267 = insertelement <16 x i8> %266, i8 %182, i64 11
+ %268 = insertelement <16 x i8> %267, i8 %184, i64 12
+ %269 = insertelement <16 x i8> %268, i8 %186, i64 13
+ %270 = insertelement <16 x i8> %269, i8 %188, i64 14
+ %271 = insertelement <16 x i8> %270, i8 %190, i64 15
+ %sum = add <16 x i8> %271, %add
+ store <16 x i8> %sum, ptr addrspace(3) %out, align 16
+ ret void
+}
+
+attributes #0 = { "amdgpu-waves-per-eu"="2,2" }
More information about the llvm-commits
mailing list