1; NOTE: Assertions have been autogenerated by utils/update_test_checks.py UTC_ARGS: --version 2 2; RUN: opt < %s -passes=loop-vectorize,instcombine -force-vector-width=2 -force-vector-interleave=1 -S | FileCheck %s 3 4; A call whose argument can remain a scalar for a vectorized function variant 5; with a uniform argument because it's loop invariant 6define void @test_uniform(ptr noalias %dst, ptr readonly %src, i64 %uniform , i64 %n) { 7; CHECK-LABEL: define void @test_uniform 8; CHECK-SAME: (ptr noalias [[DST:%.*]], ptr readonly [[SRC:%.*]], i64 [[UNIFORM:%.*]], i64 [[N:%.*]]) { 9; CHECK-NEXT: entry: 10; CHECK-NEXT: [[MIN_ITERS_CHECK:%.*]] = icmp ult i64 [[N]], 2 11; CHECK-NEXT: br i1 [[MIN_ITERS_CHECK]], label [[SCALAR_PH:%.*]], label [[VECTOR_PH:%.*]] 12; CHECK: vector.ph: 13; CHECK-NEXT: [[N_VEC:%.*]] = and i64 [[N]], -2 14; CHECK-NEXT: br label [[VECTOR_BODY:%.*]] 15; CHECK: vector.body: 16; CHECK-NEXT: [[INDEX:%.*]] = phi i64 [ 0, [[VECTOR_PH]] ], [ [[INDEX_NEXT:%.*]], [[VECTOR_BODY]] ] 17; CHECK-NEXT: [[TMP0:%.*]] = getelementptr double, ptr [[SRC]], i64 [[INDEX]] 18; CHECK-NEXT: [[WIDE_LOAD:%.*]] = load <2 x double>, ptr [[TMP0]], align 8 19; CHECK-NEXT: [[TMP1:%.*]] = call <2 x double> @foo_uniform(<2 x double> [[WIDE_LOAD]], i64 [[UNIFORM]]) 20; CHECK-NEXT: [[TMP2:%.*]] = getelementptr inbounds double, ptr [[DST]], i64 [[INDEX]] 21; CHECK-NEXT: store <2 x double> [[TMP1]], ptr [[TMP2]], align 8 22; CHECK-NEXT: [[INDEX_NEXT]] = add nuw i64 [[INDEX]], 2 23; CHECK-NEXT: [[TMP3:%.*]] = icmp eq i64 [[INDEX_NEXT]], [[N_VEC]] 24; CHECK-NEXT: br i1 [[TMP3]], label [[MIDDLE_BLOCK:%.*]], label [[VECTOR_BODY]], !llvm.loop [[LOOP0:![0-9]+]] 25; CHECK: middle.block: 26; CHECK-NEXT: [[CMP_N:%.*]] = icmp eq i64 [[N]], [[N_VEC]] 27; CHECK-NEXT: br i1 [[CMP_N]], label [[FOR_COND_CLEANUP:%.*]], label [[SCALAR_PH]] 28; CHECK: scalar.ph: 29; CHECK-NEXT: [[BC_RESUME_VAL:%.*]] = phi i64 [ [[N_VEC]], [[MIDDLE_BLOCK]] ], [ 0, [[ENTRY:%.*]] ] 30; CHECK-NEXT: br label [[FOR_BODY:%.*]] 31; CHECK: for.body: 32; CHECK-NEXT: [[INDVARS_IV:%.*]] = phi i64 [ [[BC_RESUME_VAL]], [[SCALAR_PH]] ], [ [[INDVARS_IV_NEXT:%.*]], [[FOR_BODY]] ] 33; CHECK-NEXT: [[GEPSRC:%.*]] = getelementptr double, ptr [[SRC]], i64 [[INDVARS_IV]] 34; CHECK-NEXT: [[DATA:%.*]] = load double, ptr [[GEPSRC]], align 8 35; CHECK-NEXT: [[CALL:%.*]] = call double @foo(double [[DATA]], i64 [[UNIFORM]]) #[[ATTR0:[0-9]+]] 36; CHECK-NEXT: [[GEPDST:%.*]] = getelementptr inbounds double, ptr [[DST]], i64 [[INDVARS_IV]] 37; CHECK-NEXT: store double [[CALL]], ptr [[GEPDST]], align 8 38; CHECK-NEXT: [[INDVARS_IV_NEXT]] = add nuw nsw i64 [[INDVARS_IV]], 1 39; CHECK-NEXT: [[EXITCOND:%.*]] = icmp eq i64 [[INDVARS_IV_NEXT]], [[N]] 40; CHECK-NEXT: br i1 [[EXITCOND]], label [[FOR_COND_CLEANUP]], label [[FOR_BODY]], !llvm.loop [[LOOP3:![0-9]+]] 41; CHECK: for.cond.cleanup: 42; CHECK-NEXT: ret void 43; 44entry: 45 br label %for.body 46 47for.body: 48 %indvars.iv = phi i64 [ 0, %entry ], [ %indvars.iv.next, %for.body ] 49 %gepsrc = getelementptr double, ptr %src, i64 %indvars.iv 50 %data = load double, ptr %gepsrc, align 8 51 %call = call double @foo(double %data, i64 %uniform) #0 52 %gepdst = getelementptr inbounds double, ptr %dst, i64 %indvars.iv 53 store double %call, ptr %gepdst 54 %indvars.iv.next = add nuw nsw i64 %indvars.iv, 1 55 %exitcond = icmp eq i64 %indvars.iv.next, %n 56 br i1 %exitcond, label %for.cond.cleanup, label %for.body 57 58for.cond.cleanup: 59 ret void 60} 61 62; If the parameter is not uniform, then we can't use the vector variant and 63; must fall back to scalarization. 64define void @test_uniform_not_invariant(ptr noalias %dst, ptr readonly %src, i64 %n) { 65; CHECK-LABEL: define void @test_uniform_not_invariant 66; CHECK-SAME: (ptr noalias [[DST:%.*]], ptr readonly [[SRC:%.*]], i64 [[N:%.*]]) { 67; CHECK-NEXT: entry: 68; CHECK-NEXT: [[MIN_ITERS_CHECK:%.*]] = icmp ult i64 [[N]], 2 69; CHECK-NEXT: br i1 [[MIN_ITERS_CHECK]], label [[SCALAR_PH:%.*]], label [[VECTOR_PH:%.*]] 70; CHECK: vector.ph: 71; CHECK-NEXT: [[N_VEC:%.*]] = and i64 [[N]], -2 72; CHECK-NEXT: br label [[VECTOR_BODY:%.*]] 73; CHECK: vector.body: 74; CHECK-NEXT: [[INDEX:%.*]] = phi i64 [ 0, [[VECTOR_PH]] ], [ [[INDEX_NEXT:%.*]], [[VECTOR_BODY]] ] 75; CHECK-NEXT: [[TMP0:%.*]] = or disjoint i64 [[INDEX]], 1 76; CHECK-NEXT: [[TMP1:%.*]] = getelementptr double, ptr [[SRC]], i64 [[INDEX]] 77; CHECK-NEXT: [[WIDE_LOAD:%.*]] = load <2 x double>, ptr [[TMP1]], align 8 78; CHECK-NEXT: [[TMP2:%.*]] = extractelement <2 x double> [[WIDE_LOAD]], i64 0 79; CHECK-NEXT: [[TMP3:%.*]] = call double @foo(double [[TMP2]], i64 [[INDEX]]) #[[ATTR0]] 80; CHECK-NEXT: [[TMP4:%.*]] = extractelement <2 x double> [[WIDE_LOAD]], i64 1 81; CHECK-NEXT: [[TMP5:%.*]] = call double @foo(double [[TMP4]], i64 [[TMP0]]) #[[ATTR0]] 82; CHECK-NEXT: [[TMP6:%.*]] = insertelement <2 x double> poison, double [[TMP3]], i64 0 83; CHECK-NEXT: [[TMP7:%.*]] = insertelement <2 x double> [[TMP6]], double [[TMP5]], i64 1 84; CHECK-NEXT: [[TMP8:%.*]] = getelementptr inbounds double, ptr [[DST]], i64 [[INDEX]] 85; CHECK-NEXT: store <2 x double> [[TMP7]], ptr [[TMP8]], align 8 86; CHECK-NEXT: [[INDEX_NEXT]] = add nuw i64 [[INDEX]], 2 87; CHECK-NEXT: [[TMP9:%.*]] = icmp eq i64 [[INDEX_NEXT]], [[N_VEC]] 88; CHECK-NEXT: br i1 [[TMP9]], label [[MIDDLE_BLOCK:%.*]], label [[VECTOR_BODY]], !llvm.loop [[LOOP4:![0-9]+]] 89; CHECK: middle.block: 90; CHECK-NEXT: [[CMP_N:%.*]] = icmp eq i64 [[N]], [[N_VEC]] 91; CHECK-NEXT: br i1 [[CMP_N]], label [[FOR_COND_CLEANUP:%.*]], label [[SCALAR_PH]] 92; CHECK: scalar.ph: 93; CHECK-NEXT: [[BC_RESUME_VAL:%.*]] = phi i64 [ [[N_VEC]], [[MIDDLE_BLOCK]] ], [ 0, [[ENTRY:%.*]] ] 94; CHECK-NEXT: br label [[FOR_BODY:%.*]] 95; CHECK: for.body: 96; CHECK-NEXT: [[INDVARS_IV:%.*]] = phi i64 [ [[BC_RESUME_VAL]], [[SCALAR_PH]] ], [ [[INDVARS_IV_NEXT:%.*]], [[FOR_BODY]] ] 97; CHECK-NEXT: [[GEPSRC:%.*]] = getelementptr double, ptr [[SRC]], i64 [[INDVARS_IV]] 98; CHECK-NEXT: [[DATA:%.*]] = load double, ptr [[GEPSRC]], align 8 99; CHECK-NEXT: [[CALL:%.*]] = call double @foo(double [[DATA]], i64 [[INDVARS_IV]]) #[[ATTR0]] 100; CHECK-NEXT: [[GEPDST:%.*]] = getelementptr inbounds double, ptr [[DST]], i64 [[INDVARS_IV]] 101; CHECK-NEXT: store double [[CALL]], ptr [[GEPDST]], align 8 102; CHECK-NEXT: [[INDVARS_IV_NEXT]] = add nuw nsw i64 [[INDVARS_IV]], 1 103; CHECK-NEXT: [[EXITCOND:%.*]] = icmp eq i64 [[INDVARS_IV_NEXT]], [[N]] 104; CHECK-NEXT: br i1 [[EXITCOND]], label [[FOR_COND_CLEANUP]], label [[FOR_BODY]], !llvm.loop [[LOOP5:![0-9]+]] 105; CHECK: for.cond.cleanup: 106; CHECK-NEXT: ret void 107; 108entry: 109 br label %for.body 110 111for.body: 112 %indvars.iv = phi i64 [ 0, %entry ], [ %indvars.iv.next, %for.body ] 113 %gepsrc = getelementptr double, ptr %src, i64 %indvars.iv 114 %data = load double, ptr %gepsrc, align 8 115 %call = call double @foo(double %data, i64 %indvars.iv) #0 116 %gepdst = getelementptr inbounds double, ptr %dst, i64 %indvars.iv 117 store double %call, ptr %gepdst 118 %indvars.iv.next = add nuw nsw i64 %indvars.iv, 1 119 %exitcond = icmp eq i64 %indvars.iv.next, %n 120 br i1 %exitcond, label %for.cond.cleanup, label %for.body 121 122for.cond.cleanup: 123 ret void 124} 125 126; Scalar functions 127declare double @foo(double, i64) 128 129; Vector variants 130declare <2 x double> @foo_uniform(<2 x double>, i64) 131 132; Mappings 133attributes #0 = { nounwind "vector-function-abi-variant"="_ZGV_LLVM_N2vu_foo(foo_uniform)" } 134