1; NOTE: Assertions have been autogenerated by utils/update_test_checks.py 2; RUN: opt -mtriple=amdgcn-amd-amdhsa -mcpu=gfx900 < %s -passes=loop-vectorize,dce,instcombine -S | FileCheck -check-prefix=GFX9 %s 3; RUN: opt -mtriple=amdgcn-amd-amdhsa -mcpu=fiji < %s -passes=loop-vectorize,dce,instcombine -S | FileCheck -check-prefix=VI %s 4; RUN: opt -mtriple=amdgcn-amd-amdhsa -mcpu=hawaii < %s -passes=loop-vectorize,dce,instcombine -S | FileCheck -check-prefix=CI %s 5 6define half @vectorize_v2f16_loop(ptr addrspace(1) noalias %s) { 7; GFX9-LABEL: @vectorize_v2f16_loop( 8; GFX9-NEXT: entry: 9; GFX9-NEXT: br i1 false, label [[SCALAR_PH:%.*]], label [[VECTOR_PH:%.*]] 10; GFX9: vector.ph: 11; GFX9-NEXT: br label [[VECTOR_BODY:%.*]] 12; GFX9: vector.body: 13; GFX9-NEXT: [[INDEX:%.*]] = phi i64 [ 0, [[VECTOR_PH]] ], [ [[INDEX_NEXT:%.*]], [[VECTOR_BODY]] ] 14; GFX9-NEXT: [[VEC_PHI:%.*]] = phi <2 x half> [ zeroinitializer, [[VECTOR_PH]] ], [ [[TMP2:%.*]], [[VECTOR_BODY]] ] 15; GFX9-NEXT: [[VEC_PHI1:%.*]] = phi <2 x half> [ zeroinitializer, [[VECTOR_PH]] ], [ [[TMP3:%.*]], [[VECTOR_BODY]] ] 16; GFX9-NEXT: [[TMP0:%.*]] = getelementptr inbounds half, ptr addrspace(1) [[S:%.*]], i64 [[INDEX]] 17; GFX9-NEXT: [[TMP1:%.*]] = getelementptr inbounds nuw i8, ptr addrspace(1) [[TMP0]], i64 4 18; GFX9-NEXT: [[WIDE_LOAD:%.*]] = load <2 x half>, ptr addrspace(1) [[TMP0]], align 2 19; GFX9-NEXT: [[WIDE_LOAD2:%.*]] = load <2 x half>, ptr addrspace(1) [[TMP1]], align 2 20; GFX9-NEXT: [[TMP2]] = fadd fast <2 x half> [[VEC_PHI]], [[WIDE_LOAD]] 21; GFX9-NEXT: [[TMP3]] = fadd fast <2 x half> [[VEC_PHI1]], [[WIDE_LOAD2]] 22; GFX9-NEXT: [[INDEX_NEXT]] = add nuw i64 [[INDEX]], 4 23; GFX9-NEXT: [[TMP4:%.*]] = icmp eq i64 [[INDEX_NEXT]], 256 24; GFX9-NEXT: br i1 [[TMP4]], label [[MIDDLE_BLOCK:%.*]], label [[VECTOR_BODY]], !llvm.loop [[LOOP0:![0-9]+]] 25; GFX9: middle.block: 26; GFX9-NEXT: [[BIN_RDX:%.*]] = fadd fast <2 x half> [[TMP3]], [[TMP2]] 27; GFX9-NEXT: [[TMP5:%.*]] = call fast half @llvm.vector.reduce.fadd.v2f16(half 0xH0000, <2 x half> [[BIN_RDX]]) 28; GFX9-NEXT: br i1 true, label [[FOR_END:%.*]], label [[SCALAR_PH]] 29; GFX9: scalar.ph: 30; GFX9-NEXT: br label [[FOR_BODY:%.*]] 31; GFX9: for.body: 32; GFX9-NEXT: br i1 poison, label [[FOR_END]], label [[FOR_BODY]], !llvm.loop [[LOOP2:![0-9]+]] 33; GFX9: for.end: 34; GFX9-NEXT: [[ADD_LCSSA:%.*]] = phi half [ poison, [[FOR_BODY]] ], [ [[TMP5]], [[MIDDLE_BLOCK]] ] 35; GFX9-NEXT: ret half [[ADD_LCSSA]] 36; 37; VI-LABEL: @vectorize_v2f16_loop( 38; VI-NEXT: entry: 39; VI-NEXT: br i1 false, label [[SCALAR_PH:%.*]], label [[VECTOR_PH:%.*]] 40; VI: vector.ph: 41; VI-NEXT: br label [[VECTOR_BODY:%.*]] 42; VI: vector.body: 43; VI-NEXT: [[INDEX:%.*]] = phi i64 [ 0, [[VECTOR_PH]] ], [ [[INDEX_NEXT:%.*]], [[VECTOR_BODY]] ] 44; VI-NEXT: [[VEC_PHI:%.*]] = phi <2 x half> [ zeroinitializer, [[VECTOR_PH]] ], [ [[TMP2:%.*]], [[VECTOR_BODY]] ] 45; VI-NEXT: [[VEC_PHI1:%.*]] = phi <2 x half> [ zeroinitializer, [[VECTOR_PH]] ], [ [[TMP3:%.*]], [[VECTOR_BODY]] ] 46; VI-NEXT: [[TMP0:%.*]] = getelementptr inbounds half, ptr addrspace(1) [[S:%.*]], i64 [[INDEX]] 47; VI-NEXT: [[TMP1:%.*]] = getelementptr inbounds nuw i8, ptr addrspace(1) [[TMP0]], i64 4 48; VI-NEXT: [[WIDE_LOAD:%.*]] = load <2 x half>, ptr addrspace(1) [[TMP0]], align 2 49; VI-NEXT: [[WIDE_LOAD2:%.*]] = load <2 x half>, ptr addrspace(1) [[TMP1]], align 2 50; VI-NEXT: [[TMP2]] = fadd fast <2 x half> [[VEC_PHI]], [[WIDE_LOAD]] 51; VI-NEXT: [[TMP3]] = fadd fast <2 x half> [[VEC_PHI1]], [[WIDE_LOAD2]] 52; VI-NEXT: [[INDEX_NEXT]] = add nuw i64 [[INDEX]], 4 53; VI-NEXT: [[TMP4:%.*]] = icmp eq i64 [[INDEX_NEXT]], 256 54; VI-NEXT: br i1 [[TMP4]], label [[MIDDLE_BLOCK:%.*]], label [[VECTOR_BODY]], !llvm.loop [[LOOP0:![0-9]+]] 55; VI: middle.block: 56; VI-NEXT: [[BIN_RDX:%.*]] = fadd fast <2 x half> [[TMP3]], [[TMP2]] 57; VI-NEXT: [[TMP5:%.*]] = call fast half @llvm.vector.reduce.fadd.v2f16(half 0xH0000, <2 x half> [[BIN_RDX]]) 58; VI-NEXT: br i1 true, label [[FOR_END:%.*]], label [[SCALAR_PH]] 59; VI: scalar.ph: 60; VI-NEXT: br label [[FOR_BODY:%.*]] 61; VI: for.body: 62; VI-NEXT: br i1 poison, label [[FOR_END]], label [[FOR_BODY]], !llvm.loop [[LOOP2:![0-9]+]] 63; VI: for.end: 64; VI-NEXT: [[ADD_LCSSA:%.*]] = phi half [ poison, [[FOR_BODY]] ], [ [[TMP5]], [[MIDDLE_BLOCK]] ] 65; VI-NEXT: ret half [[ADD_LCSSA]] 66; 67; CI-LABEL: @vectorize_v2f16_loop( 68; CI-NEXT: entry: 69; CI-NEXT: br label [[FOR_BODY:%.*]] 70; CI: for.body: 71; CI-NEXT: [[INDVARS_IV:%.*]] = phi i64 [ 0, [[ENTRY:%.*]] ], [ [[INDVARS_IV_NEXT:%.*]], [[FOR_BODY]] ] 72; CI-NEXT: [[Q_04:%.*]] = phi half [ 0xH0000, [[ENTRY]] ], [ [[ADD:%.*]], [[FOR_BODY]] ] 73; CI-NEXT: [[ARRAYIDX:%.*]] = getelementptr inbounds nuw half, ptr addrspace(1) [[S:%.*]], i64 [[INDVARS_IV]] 74; CI-NEXT: [[TMP0:%.*]] = load half, ptr addrspace(1) [[ARRAYIDX]], align 2 75; CI-NEXT: [[ADD]] = fadd fast half [[Q_04]], [[TMP0]] 76; CI-NEXT: [[INDVARS_IV_NEXT]] = add nuw nsw i64 [[INDVARS_IV]], 1 77; CI-NEXT: [[EXITCOND:%.*]] = icmp eq i64 [[INDVARS_IV_NEXT]], 256 78; CI-NEXT: br i1 [[EXITCOND]], label [[FOR_END:%.*]], label [[FOR_BODY]] 79; CI: for.end: 80; CI-NEXT: ret half [[ADD]] 81; 82entry: 83 br label %for.body 84 85for.body: 86 %indvars.iv = phi i64 [ 0, %entry ], [ %indvars.iv.next, %for.body ] 87 %q.04 = phi half [ 0.0, %entry ], [ %add, %for.body ] 88 %arrayidx = getelementptr inbounds half, ptr addrspace(1) %s, i64 %indvars.iv 89 %0 = load half, ptr addrspace(1) %arrayidx, align 2 90 %add = fadd fast half %q.04, %0 91 %indvars.iv.next = add nuw nsw i64 %indvars.iv, 1 92 %exitcond = icmp eq i64 %indvars.iv.next, 256 93 br i1 %exitcond, label %for.end, label %for.body 94 95for.end: 96 %add.lcssa = phi half [ %add, %for.body ] 97 ret half %add.lcssa 98} 99