1; NOTE: Assertions have been autogenerated by utils/update_test_checks.py 2; RUN: opt < %s -passes=loop-vectorize -mtriple aarch64-unknown-linux-gnu -mattr=+sve -epilogue-vectorization-force-VF=2 -S | FileCheck %s 3 4; 5; Strict fadd reduction with interleaving 6; 7define float @fadd_strict(ptr noalias nocapture readonly %a, i64 %n) { 8; CHECK-LABEL: @fadd_strict( 9; CHECK-NEXT: iter.check: 10; CHECK-NEXT: [[MIN_ITERS_CHECK:%.*]] = icmp ult i64 [[N:%.*]], 2 11; CHECK-NEXT: br i1 [[MIN_ITERS_CHECK]], label [[VEC_EPILOG_SCALAR_PH:%.*]], label [[VECTOR_MAIN_LOOP_ITER_CHECK:%.*]] 12; CHECK: vector.main.loop.iter.check: 13; CHECK-NEXT: [[TMP0:%.*]] = call i64 @llvm.vscale.i64() 14; CHECK-NEXT: [[TMP1:%.*]] = mul i64 [[TMP0]], 8 15; CHECK-NEXT: [[MIN_ITERS_CHECK1:%.*]] = icmp ult i64 [[N]], [[TMP1]] 16; CHECK-NEXT: br i1 [[MIN_ITERS_CHECK1]], label [[VEC_EPILOG_PH:%.*]], label [[VECTOR_PH:%.*]] 17; CHECK: vector.ph: 18; CHECK-NEXT: [[TMP2:%.*]] = call i64 @llvm.vscale.i64() 19; CHECK-NEXT: [[TMP3:%.*]] = mul i64 [[TMP2]], 8 20; CHECK-NEXT: [[N_MOD_VF:%.*]] = urem i64 [[N]], [[TMP3]] 21; CHECK-NEXT: [[N_VEC:%.*]] = sub i64 [[N]], [[N_MOD_VF]] 22; CHECK-NEXT: [[TMP4:%.*]] = call i64 @llvm.vscale.i64() 23; CHECK-NEXT: [[TMP5:%.*]] = mul i64 [[TMP4]], 8 24; CHECK-NEXT: br label [[VECTOR_BODY:%.*]] 25; CHECK: vector.body: 26; CHECK-NEXT: [[INDEX:%.*]] = phi i64 [ 0, [[VECTOR_PH]] ], [ [[INDEX_NEXT:%.*]], [[VECTOR_BODY]] ] 27; CHECK-NEXT: [[VEC_PHI:%.*]] = phi float [ 0xFFFFFFFFE0000000, [[VECTOR_PH]] ], [ [[TMP19:%.*]], [[VECTOR_BODY]] ] 28; CHECK-NEXT: [[TMP6:%.*]] = add i64 [[INDEX]], 0 29; CHECK-NEXT: [[TMP12:%.*]] = getelementptr inbounds float, ptr [[A:%.*]], i64 [[TMP6]] 30; CHECK-NEXT: [[TMP14:%.*]] = getelementptr inbounds float, ptr [[TMP12]], i32 0 31; CHECK-NEXT: [[TMP15:%.*]] = call i64 @llvm.vscale.i64() 32; CHECK-NEXT: [[TMP16:%.*]] = mul i64 [[TMP15]], 4 33; CHECK-NEXT: [[TMP17:%.*]] = getelementptr inbounds float, ptr [[TMP12]], i64 [[TMP16]] 34; CHECK-NEXT: [[WIDE_LOAD:%.*]] = load <vscale x 4 x float>, ptr [[TMP14]], align 4 35; CHECK-NEXT: [[WIDE_LOAD2:%.*]] = load <vscale x 4 x float>, ptr [[TMP17]], align 4 36; CHECK-NEXT: [[TMP18:%.*]] = call float @llvm.vector.reduce.fadd.nxv4f32(float [[VEC_PHI]], <vscale x 4 x float> [[WIDE_LOAD]]) 37; CHECK-NEXT: [[TMP19]] = call float @llvm.vector.reduce.fadd.nxv4f32(float [[TMP18]], <vscale x 4 x float> [[WIDE_LOAD2]]) 38; CHECK-NEXT: [[INDEX_NEXT]] = add nuw i64 [[INDEX]], [[TMP5]] 39; CHECK-NEXT: [[TMP20:%.*]] = icmp eq i64 [[INDEX_NEXT]], [[N_VEC]] 40; CHECK-NEXT: br i1 [[TMP20]], label [[MIDDLE_BLOCK:%.*]], label [[VECTOR_BODY]], !llvm.loop [[LOOP0:![0-9]+]] 41; CHECK: middle.block: 42; CHECK-NEXT: [[CMP_N:%.*]] = icmp eq i64 [[N]], [[N_VEC]] 43; CHECK-NEXT: br i1 [[CMP_N]], label [[FOR_END:%.*]], label [[VEC_EPILOG_ITER_CHECK:%.*]] 44; CHECK: vec.epilog.iter.check: 45; CHECK-NEXT: [[N_VEC_REMAINING:%.*]] = sub i64 [[N]], [[N_VEC]] 46; CHECK-NEXT: [[MIN_EPILOG_ITERS_CHECK:%.*]] = icmp ult i64 [[N_VEC_REMAINING]], 2 47; CHECK-NEXT: br i1 [[MIN_EPILOG_ITERS_CHECK]], label [[VEC_EPILOG_SCALAR_PH]], label [[VEC_EPILOG_PH]] 48; CHECK: vec.epilog.ph: 49; CHECK-NEXT: [[VEC_EPILOG_RESUME_VAL:%.*]] = phi i64 [ [[N_VEC]], [[VEC_EPILOG_ITER_CHECK]] ], [ 0, [[VECTOR_MAIN_LOOP_ITER_CHECK]] ] 50; CHECK-NEXT: [[BC_MERGE_RDX:%.*]] = phi float [ [[TMP19]], [[VEC_EPILOG_ITER_CHECK]] ], [ 0xFFFFFFFFE0000000, [[VECTOR_MAIN_LOOP_ITER_CHECK]] ] 51; CHECK-NEXT: [[N_MOD_VF3:%.*]] = urem i64 [[N]], 2 52; CHECK-NEXT: [[N_VEC4:%.*]] = sub i64 [[N]], [[N_MOD_VF3]] 53; CHECK-NEXT: br label [[VEC_EPILOG_VECTOR_BODY:%.*]] 54; CHECK: vec.epilog.vector.body: 55; CHECK-NEXT: [[INDEX6:%.*]] = phi i64 [ [[VEC_EPILOG_RESUME_VAL]], [[VEC_EPILOG_PH]] ], [ [[INDEX_NEXT9:%.*]], [[VEC_EPILOG_VECTOR_BODY]] ] 56; CHECK-NEXT: [[VEC_PHI7:%.*]] = phi float [ [[BC_MERGE_RDX]], [[VEC_EPILOG_PH]] ], [ [[TMP24:%.*]], [[VEC_EPILOG_VECTOR_BODY]] ] 57; CHECK-NEXT: [[TMP21:%.*]] = add i64 [[INDEX6]], 0 58; CHECK-NEXT: [[TMP22:%.*]] = getelementptr inbounds float, ptr [[A]], i64 [[TMP21]] 59; CHECK-NEXT: [[TMP23:%.*]] = getelementptr inbounds float, ptr [[TMP22]], i32 0 60; CHECK-NEXT: [[WIDE_LOAD8:%.*]] = load <2 x float>, ptr [[TMP23]], align 4 61; CHECK-NEXT: [[TMP24]] = call float @llvm.vector.reduce.fadd.v2f32(float [[VEC_PHI7]], <2 x float> [[WIDE_LOAD8]]) 62; CHECK-NEXT: [[INDEX_NEXT9]] = add nuw i64 [[INDEX6]], 2 63; CHECK-NEXT: [[TMP25:%.*]] = icmp eq i64 [[INDEX_NEXT9]], [[N_VEC4]] 64; CHECK-NEXT: br i1 [[TMP25]], label [[VEC_EPILOG_MIDDLE_BLOCK:%.*]], label [[VEC_EPILOG_VECTOR_BODY]], !llvm.loop [[LOOP3:![0-9]+]] 65; CHECK: vec.epilog.middle.block: 66; CHECK-NEXT: [[CMP_N5:%.*]] = icmp eq i64 [[N]], [[N_VEC4]] 67; CHECK-NEXT: br i1 [[CMP_N5]], label [[FOR_END]], label [[VEC_EPILOG_SCALAR_PH]] 68; CHECK: vec.epilog.scalar.ph: 69; CHECK-NEXT: [[BC_RESUME_VAL:%.*]] = phi i64 [ [[N_VEC4]], [[VEC_EPILOG_MIDDLE_BLOCK]] ], [ 0, [[ITER_CHECK:%.*]] ], [ [[N_VEC]], [[VEC_EPILOG_ITER_CHECK]] ] 70; CHECK-NEXT: [[BC_MERGE_RDX10:%.*]] = phi float [ [[TMP24]], [[VEC_EPILOG_MIDDLE_BLOCK]] ], [ 0xFFFFFFFFE0000000, [[ITER_CHECK]] ], [ [[TMP19]], [[VEC_EPILOG_ITER_CHECK]] ] 71; CHECK-NEXT: br label [[FOR_BODY:%.*]] 72; CHECK: for.body: 73; CHECK-NEXT: [[IV:%.*]] = phi i64 [ [[BC_RESUME_VAL]], [[VEC_EPILOG_SCALAR_PH]] ], [ [[IV_NEXT:%.*]], [[FOR_BODY]] ] 74; CHECK-NEXT: [[SUM_07:%.*]] = phi float [ [[BC_MERGE_RDX10]], [[VEC_EPILOG_SCALAR_PH]] ], [ [[ADD:%.*]], [[FOR_BODY]] ] 75; CHECK-NEXT: [[ARRAYIDX:%.*]] = getelementptr inbounds float, ptr [[A]], i64 [[IV]] 76; CHECK-NEXT: [[TMP26:%.*]] = load float, ptr [[ARRAYIDX]], align 4 77; CHECK-NEXT: [[ADD]] = fadd float [[TMP26]], [[SUM_07]] 78; CHECK-NEXT: [[IV_NEXT]] = add nuw nsw i64 [[IV]], 1 79; CHECK-NEXT: [[EXITCOND_NOT:%.*]] = icmp eq i64 [[IV_NEXT]], [[N]] 80; CHECK-NEXT: br i1 [[EXITCOND_NOT]], label [[FOR_END]], label [[FOR_BODY]], !llvm.loop [[LOOP4:![0-9]+]] 81; CHECK: for.end: 82; CHECK-NEXT: [[ADD_LCSSA:%.*]] = phi float [ [[ADD]], [[FOR_BODY]] ], [ [[TMP19]], [[MIDDLE_BLOCK]] ], [ [[TMP24]], [[VEC_EPILOG_MIDDLE_BLOCK]] ] 83; CHECK-NEXT: ret float [[ADD_LCSSA]] 84; 85entry: 86 br label %for.body 87 88for.body: 89 %iv = phi i64 [ 0, %entry ], [ %iv.next, %for.body ] 90 %sum.07 = phi float [ 0xFFFFFFFFE0000000, %entry ], [ %add, %for.body ] 91 %arrayidx = getelementptr inbounds float, ptr %a, i64 %iv 92 %0 = load float, ptr %arrayidx, align 4 93 %add = fadd float %0, %sum.07 94 %iv.next = add nuw nsw i64 %iv, 1 95 %exitcond.not = icmp eq i64 %iv.next, %n 96 br i1 %exitcond.not, label %for.end, label %for.body, !llvm.loop !0 97 98for.end: 99 ret float %add 100} 101 102!0 = distinct !{!0, !1, !2} 103!1 = !{!"llvm.loop.interleave.count", i32 2} 104!2 = !{!"llvm.loop.vectorize.scalable.enable", i1 true} 105