106 lines · plain
1; RUN: llc < %s -mtriple=amdgcn -mcpu=verde | FileCheck %s2; RUN: llc < %s -mtriple=amdgcn -mcpu=tonga -mattr=-flat-for-global | FileCheck %s3 4; CHECK: {{^}}test_8_min_char:5; CHECK: buffer_store_byte6; CHECK: buffer_store_byte7; CHECK: buffer_store_byte8; CHECK: buffer_store_byte9; CHECK: buffer_store_byte10; CHECK: buffer_store_byte11; CHECK: buffer_store_byte12; CHECK: buffer_store_byte13; ModuleID = 'radeon'14 15define amdgpu_kernel void @test_8_min_char(ptr addrspace(1) nocapture %out, ptr addrspace(1) nocapture readonly %in0, ptr addrspace(1) nocapture readonly %in1) #0 {16entry:17 %0 = load i8, ptr addrspace(1) %in0, align 118 %1 = insertelement <8 x i8> poison, i8 %0, i32 019 %arrayidx2.i.i = getelementptr inbounds i8, ptr addrspace(1) %in0, i64 120 %2 = load i8, ptr addrspace(1) %arrayidx2.i.i, align 121 %3 = insertelement <8 x i8> %1, i8 %2, i32 122 %arrayidx6.i.i = getelementptr inbounds i8, ptr addrspace(1) %in0, i64 223 %4 = load i8, ptr addrspace(1) %arrayidx6.i.i, align 124 %5 = insertelement <8 x i8> %3, i8 %4, i32 225 %arrayidx10.i.i = getelementptr inbounds i8, ptr addrspace(1) %in0, i64 326 %6 = load i8, ptr addrspace(1) %arrayidx10.i.i, align 127 %7 = insertelement <8 x i8> %5, i8 %6, i32 328 %arrayidx.i.i = getelementptr inbounds i8, ptr addrspace(1) %in0, i64 429 %8 = load i8, ptr addrspace(1) %arrayidx.i.i, align 130 %9 = insertelement <8 x i8> poison, i8 %8, i32 031 %arrayidx2.i9.i = getelementptr inbounds i8, ptr addrspace(1) %in0, i64 532 %10 = load i8, ptr addrspace(1) %arrayidx2.i9.i, align 133 %11 = insertelement <8 x i8> %9, i8 %10, i32 134 %arrayidx6.i11.i = getelementptr inbounds i8, ptr addrspace(1) %in0, i64 635 %12 = load i8, ptr addrspace(1) %arrayidx6.i11.i, align 136 %13 = insertelement <8 x i8> %11, i8 %12, i32 237 %arrayidx10.i13.i = getelementptr inbounds i8, ptr addrspace(1) %in0, i64 738 %14 = load i8, ptr addrspace(1) %arrayidx10.i13.i, align 139 %15 = insertelement <8 x i8> %13, i8 %14, i32 340 %vecinit5.i = shufflevector <8 x i8> %7, <8 x i8> %15, <8 x i32> <i32 0, i32 1, i32 2, i32 3, i32 8, i32 9, i32 10, i32 11>41 %16 = load i8, ptr addrspace(1) %in1, align 142 %17 = insertelement <8 x i8> poison, i8 %16, i32 043 %arrayidx2.i.i4 = getelementptr inbounds i8, ptr addrspace(1) %in1, i64 144 %18 = load i8, ptr addrspace(1) %arrayidx2.i.i4, align 145 %19 = insertelement <8 x i8> %17, i8 %18, i32 146 %arrayidx6.i.i5 = getelementptr inbounds i8, ptr addrspace(1) %in1, i64 247 %20 = load i8, ptr addrspace(1) %arrayidx6.i.i5, align 148 %21 = insertelement <8 x i8> %19, i8 %20, i32 249 %arrayidx10.i.i6 = getelementptr inbounds i8, ptr addrspace(1) %in1, i64 350 %22 = load i8, ptr addrspace(1) %arrayidx10.i.i6, align 151 %23 = insertelement <8 x i8> %21, i8 %22, i32 352 %arrayidx.i.i7 = getelementptr inbounds i8, ptr addrspace(1) %in1, i64 453 %24 = load i8, ptr addrspace(1) %arrayidx.i.i7, align 154 %25 = insertelement <8 x i8> poison, i8 %24, i32 055 %arrayidx2.i9.i8 = getelementptr inbounds i8, ptr addrspace(1) %in1, i64 556 %26 = load i8, ptr addrspace(1) %arrayidx2.i9.i8, align 157 %27 = insertelement <8 x i8> %25, i8 %26, i32 158 %arrayidx6.i11.i9 = getelementptr inbounds i8, ptr addrspace(1) %in1, i64 659 %28 = load i8, ptr addrspace(1) %arrayidx6.i11.i9, align 160 %29 = insertelement <8 x i8> %27, i8 %28, i32 261 %arrayidx10.i13.i10 = getelementptr inbounds i8, ptr addrspace(1) %in1, i64 762 %30 = load i8, ptr addrspace(1) %arrayidx10.i13.i10, align 163 %31 = insertelement <8 x i8> %29, i8 %30, i32 364 %vecinit5.i11 = shufflevector <8 x i8> %23, <8 x i8> %31, <8 x i32> <i32 0, i32 1, i32 2, i32 3, i32 8, i32 9, i32 10, i32 11>65 %cmp.i = icmp slt <8 x i8> %vecinit5.i, %vecinit5.i1166 %cond.i = select <8 x i1> %cmp.i, <8 x i8> %vecinit5.i, <8 x i8> %vecinit5.i1167 %32 = extractelement <8 x i8> %cond.i, i32 068 store i8 %32, ptr addrspace(1) %out, align 169 %33 = extractelement <8 x i8> %cond.i, i32 170 %arrayidx2.i.i.i = getelementptr inbounds i8, ptr addrspace(1) %out, i64 171 store i8 %33, ptr addrspace(1) %arrayidx2.i.i.i, align 172 %34 = extractelement <8 x i8> %cond.i, i32 273 %arrayidx.i.i.i = getelementptr inbounds i8, ptr addrspace(1) %out, i64 274 store i8 %34, ptr addrspace(1) %arrayidx.i.i.i, align 175 %35 = extractelement <8 x i8> %cond.i, i32 376 %arrayidx2.i6.i.i = getelementptr inbounds i8, ptr addrspace(1) %out, i64 377 store i8 %35, ptr addrspace(1) %arrayidx2.i6.i.i, align 178 %arrayidx.i.i3 = getelementptr inbounds i8, ptr addrspace(1) %out, i64 479 %36 = extractelement <8 x i8> %cond.i, i32 480 store i8 %36, ptr addrspace(1) %arrayidx.i.i3, align 181 %37 = extractelement <8 x i8> %cond.i, i32 582 %arrayidx2.i.i6.i = getelementptr inbounds i8, ptr addrspace(1) %out, i64 583 store i8 %37, ptr addrspace(1) %arrayidx2.i.i6.i, align 184 %38 = extractelement <8 x i8> %cond.i, i32 685 %arrayidx.i.i7.i = getelementptr inbounds i8, ptr addrspace(1) %out, i64 686 store i8 %38, ptr addrspace(1) %arrayidx.i.i7.i, align 187 %39 = extractelement <8 x i8> %cond.i, i32 788 %arrayidx2.i6.i8.i = getelementptr inbounds i8, ptr addrspace(1) %out, i64 789 store i8 %39, ptr addrspace(1) %arrayidx2.i6.i8.i, align 190 ret void91}92 93attributes #0 = { nounwind }94 95!opencl.kernels = !{!0, !1, !2, !3, !4, !5, !6, !7, !8}96 97!0 = !{null}98!1 = !{null}99!2 = !{null}100!3 = !{ptr @test_8_min_char}101!4 = !{null}102!5 = !{null}103!6 = !{null}104!7 = !{null}105!8 = !{null}106