llvm/test/Transforms/LoadStoreVectorizer/NVPTX/gap-fill-with-redundant-elements.ll

; NOTE: Assertions have been autogenerated by utils/update_test_checks.py UTC_ARGS: --version 6
; RUN: opt -mtriple=nvptx64-nvidia-cuda -passes=load-store-vectorizer -mcpu=sm_100 -mattr=+ptx88 -S -o - %s | FileCheck %s

define void @test_redundant_no_gap(ptr addrspace(1) %ptr) {
; CHECK-LABEL: define void @test_redundant_no_gap(
; CHECK-SAME: ptr addrspace(1) [[PTR:%.*]]) #[[ATTR0:[0-9]+]] {
; CHECK-NEXT:    [[TMP1:%.*]] = load <8 x i32>, ptr addrspace(1) [[PTR]], align 32
; CHECK-NEXT:    [[LD03:%.*]] = shufflevector <8 x i32> [[TMP1]], <8 x i32> poison, <2 x i32> <i32 0, i32 1>
; CHECK-NEXT:    [[LD12:%.*]] = extractelement <8 x i32> [[TMP1]], i64 1
; CHECK-NEXT:    [[LD25:%.*]] = shufflevector <8 x i32> [[TMP1]], <8 x i32> poison, <2 x i32> <i32 2, i32 3>
; CHECK-NEXT:    [[LD37:%.*]] = shufflevector <8 x i32> [[TMP1]], <8 x i32> poison, <2 x i32> <i32 4, i32 5>
; CHECK-NEXT:    [[LD45:%.*]] = shufflevector <8 x i32> [[TMP1]], <8 x i32> poison, <2 x i32> <i32 6, i32 7>
; CHECK-NEXT:    ret void
;
  %ld0 = load <2 x i32>, ptr addrspace(1) %ptr, align 32
  %gep1 = getelementptr inbounds i8, ptr addrspace(1) %ptr, i32 4
  %ld1 = load i32, ptr addrspace(1) %gep1, align 4
  %gep2 = getelementptr inbounds i8, ptr addrspace(1) %ptr, i32 8
  %ld2 = load <2 x i32>, ptr addrspace(1) %gep2, align 8
  %gep3 = getelementptr inbounds i8, ptr addrspace(1) %ptr, i32 16
  %ld3 = load <2 x i32>, ptr addrspace(1) %gep3, align 16
  %gep4 = getelementptr inbounds i8, ptr addrspace(1) %ptr, i32 24
  %ld4 = load <2 x i32>, ptr addrspace(1) %gep4, align 8
  ret void
}

; This fills the 2-byte gap between ld0 and ld3.
; ld1 is folded into the vector but ld0 is treated as the end of the chain
; at the point when the gap is considered, because it reads further than ld1.
define void @test_redundant_gap(ptr addrspace(1) %ptr) {
; CHECK-LABEL: define void @test_redundant_gap(
; CHECK-SAME: ptr addrspace(1) [[PTR:%.*]]) #[[ATTR0]] {
; CHECK-NEXT:    [[TMP1:%.*]] = call <8 x i32> @llvm.masked.load.v8i32.p1(ptr addrspace(1) align 32 [[PTR]], <8 x i1> <i1 true, i1 true, i1 false, i1 false, i1 true, i1 true, i1 true, i1 true>, <8 x i32> poison)
; CHECK-NEXT:    [[LD01:%.*]] = shufflevector <8 x i32> [[TMP1]], <8 x i32> poison, <2 x i32> <i32 0, i32 1>
; CHECK-NEXT:    [[LD12:%.*]] = extractelement <8 x i32> [[TMP1]], i64 0
; CHECK-NEXT:    [[GAPFILL3:%.*]] = shufflevector <8 x i32> [[TMP1]], <8 x i32> poison, <2 x i32> <i32 2, i32 3>
; CHECK-NEXT:    [[LD34:%.*]] = shufflevector <8 x i32> [[TMP1]], <8 x i32> poison, <2 x i32> <i32 4, i32 5>
; CHECK-NEXT:    [[LD45:%.*]] = shufflevector <8 x i32> [[TMP1]], <8 x i32> poison, <2 x i32> <i32 6, i32 7>
; CHECK-NEXT:    ret void
;
  %ld0 = load <2 x i32>, ptr addrspace(1) %ptr, align 32
  %ld1 = load i32, ptr addrspace(1) %ptr, align 4
  %gep3 = getelementptr inbounds i8, ptr addrspace(1) %ptr, i32 16
  %ld3 = load <2 x i32>, ptr addrspace(1) %gep3, align 16
  %gep4 = getelementptr inbounds i8, ptr addrspace(1) %ptr, i32 24
  %ld4 = load <2 x i32>, ptr addrspace(1) %gep4, align 8
  ret void
}

; This chain contains two elements, one before a gap,
; and one before the end of the chain. Chain should be correctly extended.
define void @test_redundant_gap_and_extend(ptr addrspace(1) %ptr) {
; CHECK-LABEL: define void @test_redundant_gap_and_extend(
; CHECK-SAME: ptr addrspace(1) [[PTR:%.*]]) #[[ATTR0]] {
; CHECK-NEXT:    [[TMP1:%.*]] = call <8 x i32> @llvm.masked.load.v8i32.p1(ptr addrspace(1) align 32 [[PTR]], <8 x i1> <i1 true, i1 true, i1 false, i1 false, i1 true, i1 true, i1 false, i1 false>, <8 x i32> poison)
; CHECK-NEXT:    [[LD03:%.*]] = shufflevector <8 x i32> [[TMP1]], <8 x i32> poison, <2 x i32> <i32 0, i32 1>
; CHECK-NEXT:    [[LD14:%.*]] = extractelement <8 x i32> [[TMP1]], i64 0
; CHECK-NEXT:    [[GAPFILL5:%.*]] = shufflevector <8 x i32> [[TMP1]], <8 x i32> poison, <2 x i32> <i32 2, i32 3>
; CHECK-NEXT:    [[LD36:%.*]] = shufflevector <8 x i32> [[TMP1]], <8 x i32> poison, <2 x i32> <i32 4, i32 5>
; CHECK-NEXT:    [[LD47:%.*]] = extractelement <8 x i32> [[TMP1]], i64 4
; CHECK-NEXT:    [[EXTEND8:%.*]] = extractelement <8 x i32> [[TMP1]], i64 6
; CHECK-NEXT:    [[EXTEND29:%.*]] = extractelement <8 x i32> [[TMP1]], i64 7
; CHECK-NEXT:    ret void
;
  %ld0 = load <2 x i32>, ptr addrspace(1) %ptr, align 32
  %ld1 = load i32, ptr addrspace(1) %ptr, align 4
  %gep3 = getelementptr inbounds i8, ptr addrspace(1) %ptr, i32 16
  %ld3 = load <2 x i32>, ptr addrspace(1) %gep3, align 16
  %ld4 = load i32, ptr addrspace(1) %gep3, align 4
  ret void
}