342 lines · plain
1; NOTE: Assertions have been autogenerated by utils/update_test_checks.py2; RUN: opt -passes=loop-distribute -enable-loop-distribute -verify-loop-info -verify-dom-info -S -enable-mem-access-versioning=0 < %s | FileCheck %s3 4; PredicatedScalarEvolution decides it needs to insert a bounds check5; not based on memory access.6 7define void @f(ptr noalias %a, ptr noalias %b, ptr noalias %c, ptr noalias %d, ptr noalias %e, i64 %N) {8; CHECK-LABEL: @f(9; CHECK-NEXT: entry:10; CHECK-NEXT: br label [[FOR_BODY_LVER_CHECK:%.*]]11; CHECK: for.body.lver.check:12; CHECK-NEXT: [[TMP0:%.*]] = add i64 [[N:%.*]], -113; CHECK-NEXT: [[TMP2:%.*]] = trunc i64 [[TMP0]] to i3214; CHECK-NEXT: [[MUL1:%.*]] = call { i32, i1 } @llvm.umul.with.overflow.i32(i32 2, i32 [[TMP2]])15; CHECK-NEXT: [[MUL_OVERFLOW:%.*]] = extractvalue { i32, i1 } [[MUL1]], 116; CHECK-NEXT: [[TMP1:%.*]] = icmp ugt i64 [[TMP0]], 429496729517; CHECK-NEXT: [[TMP3:%.*]] = or i1 [[MUL_OVERFLOW]], [[TMP1]]18; CHECK-NEXT: br i1 [[TMP3]], label [[FOR_BODY_PH_LVER_ORIG:%.*]], label [[FOR_BODY_PH_LDIST1:%.*]]19; CHECK: for.body.ph.lver.orig:20; CHECK-NEXT: br label [[FOR_BODY_LVER_ORIG:%.*]]21; CHECK: for.body.lver.orig:22; CHECK-NEXT: [[IND_LVER_ORIG:%.*]] = phi i64 [ 0, [[FOR_BODY_PH_LVER_ORIG]] ], [ [[ADD_LVER_ORIG:%.*]], [[FOR_BODY_LVER_ORIG]] ]23; CHECK-NEXT: [[IND1_LVER_ORIG:%.*]] = phi i32 [ 0, [[FOR_BODY_PH_LVER_ORIG]] ], [ [[INC1_LVER_ORIG:%.*]], [[FOR_BODY_LVER_ORIG]] ]24; CHECK-NEXT: [[MUL_LVER_ORIG:%.*]] = mul i32 [[IND1_LVER_ORIG]], 225; CHECK-NEXT: [[MUL_EXT_LVER_ORIG:%.*]] = zext i32 [[MUL_LVER_ORIG]] to i6426; CHECK-NEXT: [[ARRAYIDXA_LVER_ORIG:%.*]] = getelementptr inbounds i32, ptr [[A:%.*]], i64 [[MUL_EXT_LVER_ORIG]]27; CHECK-NEXT: [[LOADA_LVER_ORIG:%.*]] = load i32, ptr [[ARRAYIDXA_LVER_ORIG]], align 428; CHECK-NEXT: [[ARRAYIDXB_LVER_ORIG:%.*]] = getelementptr inbounds i32, ptr [[B:%.*]], i64 [[MUL_EXT_LVER_ORIG]]29; CHECK-NEXT: [[LOADB_LVER_ORIG:%.*]] = load i32, ptr [[ARRAYIDXB_LVER_ORIG]], align 430; CHECK-NEXT: [[MULA_LVER_ORIG:%.*]] = mul i32 [[LOADB_LVER_ORIG]], [[LOADA_LVER_ORIG]]31; CHECK-NEXT: [[ADD_LVER_ORIG]] = add nuw nsw i64 [[IND_LVER_ORIG]], 132; CHECK-NEXT: [[INC1_LVER_ORIG]] = add i32 [[IND1_LVER_ORIG]], 133; CHECK-NEXT: [[ARRAYIDXA_PLUS_4_LVER_ORIG:%.*]] = getelementptr inbounds i32, ptr [[A]], i64 [[ADD_LVER_ORIG]]34; CHECK-NEXT: store i32 [[MULA_LVER_ORIG]], ptr [[ARRAYIDXA_PLUS_4_LVER_ORIG]], align 435; CHECK-NEXT: [[ARRAYIDXD_LVER_ORIG:%.*]] = getelementptr inbounds i32, ptr [[D:%.*]], i64 [[MUL_EXT_LVER_ORIG]]36; CHECK-NEXT: [[LOADD_LVER_ORIG:%.*]] = load i32, ptr [[ARRAYIDXD_LVER_ORIG]], align 437; CHECK-NEXT: [[ARRAYIDXE_LVER_ORIG:%.*]] = getelementptr inbounds i32, ptr [[E:%.*]], i64 [[MUL_EXT_LVER_ORIG]]38; CHECK-NEXT: [[LOADE_LVER_ORIG:%.*]] = load i32, ptr [[ARRAYIDXE_LVER_ORIG]], align 439; CHECK-NEXT: [[MULC_LVER_ORIG:%.*]] = mul i32 [[LOADD_LVER_ORIG]], [[LOADE_LVER_ORIG]]40; CHECK-NEXT: [[ARRAYIDXC_LVER_ORIG:%.*]] = getelementptr inbounds i32, ptr [[C:%.*]], i64 [[MUL_EXT_LVER_ORIG]]41; CHECK-NEXT: store i32 [[MULC_LVER_ORIG]], ptr [[ARRAYIDXC_LVER_ORIG]], align 442; CHECK-NEXT: [[EXITCOND_LVER_ORIG:%.*]] = icmp eq i64 [[ADD_LVER_ORIG]], [[N]]43; CHECK-NEXT: br i1 [[EXITCOND_LVER_ORIG]], label [[FOR_END_LOOPEXIT:%.*]], label [[FOR_BODY_LVER_ORIG]], !llvm.loop [[LOOP0:![0-9]+]]44; CHECK: for.body.ph.ldist1:45; CHECK-NEXT: br label [[FOR_BODY_LDIST1:%.*]]46; CHECK: for.body.ldist1:47; CHECK-NEXT: [[IND_LDIST1:%.*]] = phi i64 [ 0, [[FOR_BODY_PH_LDIST1]] ], [ [[ADD_LDIST1:%.*]], [[FOR_BODY_LDIST1]] ]48; CHECK-NEXT: [[IND1_LDIST1:%.*]] = phi i32 [ 0, [[FOR_BODY_PH_LDIST1]] ], [ [[INC1_LDIST1:%.*]], [[FOR_BODY_LDIST1]] ]49; CHECK-NEXT: [[MUL_LDIST1:%.*]] = mul i32 [[IND1_LDIST1]], 250; CHECK-NEXT: [[MUL_EXT_LDIST1:%.*]] = zext i32 [[MUL_LDIST1]] to i6451; CHECK-NEXT: [[ARRAYIDXA_LDIST1:%.*]] = getelementptr inbounds i32, ptr [[A]], i64 [[MUL_EXT_LDIST1]]52; CHECK-NEXT: [[LOADA_LDIST1:%.*]] = load i32, ptr [[ARRAYIDXA_LDIST1]], align 4, !alias.scope [[META2:![0-9]+]]53; CHECK-NEXT: [[ARRAYIDXB_LDIST1:%.*]] = getelementptr inbounds i32, ptr [[B]], i64 [[MUL_EXT_LDIST1]]54; CHECK-NEXT: [[LOADB_LDIST1:%.*]] = load i32, ptr [[ARRAYIDXB_LDIST1]], align 455; CHECK-NEXT: [[MULA_LDIST1:%.*]] = mul i32 [[LOADB_LDIST1]], [[LOADA_LDIST1]]56; CHECK-NEXT: [[ADD_LDIST1]] = add nuw nsw i64 [[IND_LDIST1]], 157; CHECK-NEXT: [[INC1_LDIST1]] = add i32 [[IND1_LDIST1]], 158; CHECK-NEXT: [[ARRAYIDXA_PLUS_4_LDIST1:%.*]] = getelementptr inbounds i32, ptr [[A]], i64 [[ADD_LDIST1]]59; CHECK-NEXT: store i32 [[MULA_LDIST1]], ptr [[ARRAYIDXA_PLUS_4_LDIST1]], align 4, !alias.scope [[META5:![0-9]+]]60; CHECK-NEXT: [[EXITCOND_LDIST1:%.*]] = icmp eq i64 [[ADD_LDIST1]], [[N]]61; CHECK-NEXT: br i1 [[EXITCOND_LDIST1]], label [[FOR_BODY_PH:%.*]], label [[FOR_BODY_LDIST1]]62; CHECK: for.body.ph:63; CHECK-NEXT: br label [[FOR_BODY:%.*]]64; CHECK: for.body:65; CHECK-NEXT: [[IND:%.*]] = phi i64 [ 0, [[FOR_BODY_PH]] ], [ [[ADD:%.*]], [[FOR_BODY]] ]66; CHECK-NEXT: [[IND1:%.*]] = phi i32 [ 0, [[FOR_BODY_PH]] ], [ [[INC1:%.*]], [[FOR_BODY]] ]67; CHECK-NEXT: [[MUL:%.*]] = mul i32 [[IND1]], 268; CHECK-NEXT: [[MUL_EXT:%.*]] = zext i32 [[MUL]] to i6469; CHECK-NEXT: [[ADD]] = add nuw nsw i64 [[IND]], 170; CHECK-NEXT: [[INC1]] = add i32 [[IND1]], 171; CHECK-NEXT: [[ARRAYIDXD:%.*]] = getelementptr inbounds i32, ptr [[D]], i64 [[MUL_EXT]]72; CHECK-NEXT: [[LOADD:%.*]] = load i32, ptr [[ARRAYIDXD]], align 473; CHECK-NEXT: [[ARRAYIDXE:%.*]] = getelementptr inbounds i32, ptr [[E]], i64 [[MUL_EXT]]74; CHECK-NEXT: [[LOADE:%.*]] = load i32, ptr [[ARRAYIDXE]], align 475; CHECK-NEXT: [[MULC:%.*]] = mul i32 [[LOADD]], [[LOADE]]76; CHECK-NEXT: [[ARRAYIDXC:%.*]] = getelementptr inbounds i32, ptr [[C]], i64 [[MUL_EXT]]77; CHECK-NEXT: store i32 [[MULC]], ptr [[ARRAYIDXC]], align 478; CHECK-NEXT: [[EXITCOND:%.*]] = icmp eq i64 [[ADD]], [[N]]79; CHECK-NEXT: br i1 [[EXITCOND]], label [[FOR_END_LOOPEXIT2:%.*]], label [[FOR_BODY]]80; CHECK: for.end.loopexit:81; CHECK-NEXT: br label [[FOR_END:%.*]]82; CHECK: for.end.loopexit2:83; CHECK-NEXT: br label [[FOR_END]]84; CHECK: for.end:85; CHECK-NEXT: ret void86;87entry:88 br label %for.body89 90for.body: ; preds = %for.body, %entry91 %ind = phi i64 [ 0, %entry ], [ %add, %for.body ]92 %ind1 = phi i32 [ 0, %entry ], [ %inc1, %for.body ]93 94 %mul = mul i32 %ind1, 295 %mul_ext = zext i32 %mul to i6496 97 98 %arrayidxA = getelementptr inbounds i32, ptr %a, i64 %mul_ext99 %loadA = load i32, ptr %arrayidxA, align 4100 101 %arrayidxB = getelementptr inbounds i32, ptr %b, i64 %mul_ext102 %loadB = load i32, ptr %arrayidxB, align 4103 104 %mulA = mul i32 %loadB, %loadA105 106 %add = add nuw nsw i64 %ind, 1107 %inc1 = add i32 %ind1, 1108 109 %arrayidxA_plus_4 = getelementptr inbounds i32, ptr %a, i64 %add110 store i32 %mulA, ptr %arrayidxA_plus_4, align 4111 112 %arrayidxD = getelementptr inbounds i32, ptr %d, i64 %mul_ext113 %loadD = load i32, ptr %arrayidxD, align 4114 115 %arrayidxE = getelementptr inbounds i32, ptr %e, i64 %mul_ext116 %loadE = load i32, ptr %arrayidxE, align 4117 118 %mulC = mul i32 %loadD, %loadE119 120 %arrayidxC = getelementptr inbounds i32, ptr %c, i64 %mul_ext121 store i32 %mulC, ptr %arrayidxC, align 4122 123 %exitcond = icmp eq i64 %add, %N124 br i1 %exitcond, label %for.end, label %for.body125 126for.end: ; preds = %for.body127 ret void128}129 130declare void @use64(i64)131@global_a = common local_unnamed_addr global [8192 x i32] zeroinitializer, align 16132 133define void @f_with_offset(ptr noalias %b, ptr noalias %c, ptr noalias %d, ptr noalias %e, i64 %N) {134; CHECK-LABEL: @f_with_offset(135; CHECK-NEXT: entry:136; CHECK-NEXT: [[A_INTPTR:%.*]] = ptrtoint ptr @global_a to i64137; CHECK-NEXT: call void @use64(i64 [[A_INTPTR]])138; CHECK-NEXT: [[A:%.*]] = getelementptr i32, ptr @global_a, i32 42139; CHECK-NEXT: br label [[FOR_BODY_LVER_CHECK:%.*]]140; CHECK: for.body.lver.check:141; CHECK-NEXT: [[TMP0:%.*]] = add i64 [[N:%.*]], -1142; CHECK-NEXT: [[TMP2:%.*]] = trunc i64 [[TMP0]] to i32143; CHECK-NEXT: [[MUL1:%.*]] = call { i32, i1 } @llvm.umul.with.overflow.i32(i32 2, i32 [[TMP2]])144; CHECK-NEXT: [[MUL_OVERFLOW:%.*]] = extractvalue { i32, i1 } [[MUL1]], 1145; CHECK-NEXT: [[TMP1:%.*]] = icmp ugt i64 [[TMP0]], 4294967295146; CHECK-NEXT: [[TMP3:%.*]] = or i1 [[MUL_OVERFLOW]], [[TMP1]]147; CHECK-NEXT: br i1 [[TMP3]], label [[FOR_BODY_PH_LVER_ORIG:%.*]], label [[FOR_BODY_PH_LDIST1:%.*]]148; CHECK: for.body.ph.lver.orig:149; CHECK-NEXT: br label [[FOR_BODY_LVER_ORIG:%.*]]150; CHECK: for.body.lver.orig:151; CHECK-NEXT: [[IND_LVER_ORIG:%.*]] = phi i64 [ 0, [[FOR_BODY_PH_LVER_ORIG]] ], [ [[ADD_LVER_ORIG:%.*]], [[FOR_BODY_LVER_ORIG]] ]152; CHECK-NEXT: [[IND1_LVER_ORIG:%.*]] = phi i32 [ 0, [[FOR_BODY_PH_LVER_ORIG]] ], [ [[INC1_LVER_ORIG:%.*]], [[FOR_BODY_LVER_ORIG]] ]153; CHECK-NEXT: [[MUL_LVER_ORIG:%.*]] = mul i32 [[IND1_LVER_ORIG]], 2154; CHECK-NEXT: [[MUL_EXT_LVER_ORIG:%.*]] = zext i32 [[MUL_LVER_ORIG]] to i64155; CHECK-NEXT: [[ARRAYIDXA_LVER_ORIG:%.*]] = getelementptr inbounds i32, ptr [[A]], i64 [[MUL_EXT_LVER_ORIG]]156; CHECK-NEXT: [[LOADA_LVER_ORIG:%.*]] = load i32, ptr [[ARRAYIDXA_LVER_ORIG]], align 4157; CHECK-NEXT: [[ARRAYIDXB_LVER_ORIG:%.*]] = getelementptr inbounds i32, ptr [[B:%.*]], i64 [[MUL_EXT_LVER_ORIG]]158; CHECK-NEXT: [[LOADB_LVER_ORIG:%.*]] = load i32, ptr [[ARRAYIDXB_LVER_ORIG]], align 4159; CHECK-NEXT: [[MULA_LVER_ORIG:%.*]] = mul i32 [[LOADB_LVER_ORIG]], [[LOADA_LVER_ORIG]]160; CHECK-NEXT: [[ADD_LVER_ORIG]] = add nuw nsw i64 [[IND_LVER_ORIG]], 1161; CHECK-NEXT: [[INC1_LVER_ORIG]] = add i32 [[IND1_LVER_ORIG]], 1162; CHECK-NEXT: [[ARRAYIDXA_PLUS_4_LVER_ORIG:%.*]] = getelementptr inbounds i32, ptr [[A]], i64 [[ADD_LVER_ORIG]]163; CHECK-NEXT: store i32 [[MULA_LVER_ORIG]], ptr [[ARRAYIDXA_PLUS_4_LVER_ORIG]], align 4164; CHECK-NEXT: [[ARRAYIDXD_LVER_ORIG:%.*]] = getelementptr inbounds i32, ptr [[D:%.*]], i64 [[MUL_EXT_LVER_ORIG]]165; CHECK-NEXT: [[LOADD_LVER_ORIG:%.*]] = load i32, ptr [[ARRAYIDXD_LVER_ORIG]], align 4166; CHECK-NEXT: [[ARRAYIDXE_LVER_ORIG:%.*]] = getelementptr inbounds i32, ptr [[E:%.*]], i64 [[MUL_EXT_LVER_ORIG]]167; CHECK-NEXT: [[LOADE_LVER_ORIG:%.*]] = load i32, ptr [[ARRAYIDXE_LVER_ORIG]], align 4168; CHECK-NEXT: [[MULC_LVER_ORIG:%.*]] = mul i32 [[LOADD_LVER_ORIG]], [[LOADE_LVER_ORIG]]169; CHECK-NEXT: [[ARRAYIDXC_LVER_ORIG:%.*]] = getelementptr inbounds i32, ptr [[C:%.*]], i64 [[MUL_EXT_LVER_ORIG]]170; CHECK-NEXT: store i32 [[MULC_LVER_ORIG]], ptr [[ARRAYIDXC_LVER_ORIG]], align 4171; CHECK-NEXT: [[EXITCOND_LVER_ORIG:%.*]] = icmp eq i64 [[ADD_LVER_ORIG]], [[N]]172; CHECK-NEXT: br i1 [[EXITCOND_LVER_ORIG]], label [[FOR_END_LOOPEXIT:%.*]], label [[FOR_BODY_LVER_ORIG]], !llvm.loop [[LOOP7:![0-9]+]]173; CHECK: for.body.ph.ldist1:174; CHECK-NEXT: br label [[FOR_BODY_LDIST1:%.*]]175; CHECK: for.body.ldist1:176; CHECK-NEXT: [[IND_LDIST1:%.*]] = phi i64 [ 0, [[FOR_BODY_PH_LDIST1]] ], [ [[ADD_LDIST1:%.*]], [[FOR_BODY_LDIST1]] ]177; CHECK-NEXT: [[IND1_LDIST1:%.*]] = phi i32 [ 0, [[FOR_BODY_PH_LDIST1]] ], [ [[INC1_LDIST1:%.*]], [[FOR_BODY_LDIST1]] ]178; CHECK-NEXT: [[MUL_LDIST1:%.*]] = mul i32 [[IND1_LDIST1]], 2179; CHECK-NEXT: [[MUL_EXT_LDIST1:%.*]] = zext i32 [[MUL_LDIST1]] to i64180; CHECK-NEXT: [[ARRAYIDXA_LDIST1:%.*]] = getelementptr inbounds i32, ptr [[A]], i64 [[MUL_EXT_LDIST1]]181; CHECK-NEXT: [[LOADA_LDIST1:%.*]] = load i32, ptr [[ARRAYIDXA_LDIST1]], align 4, !alias.scope [[META8:![0-9]+]]182; CHECK-NEXT: [[ARRAYIDXB_LDIST1:%.*]] = getelementptr inbounds i32, ptr [[B]], i64 [[MUL_EXT_LDIST1]]183; CHECK-NEXT: [[LOADB_LDIST1:%.*]] = load i32, ptr [[ARRAYIDXB_LDIST1]], align 4184; CHECK-NEXT: [[MULA_LDIST1:%.*]] = mul i32 [[LOADB_LDIST1]], [[LOADA_LDIST1]]185; CHECK-NEXT: [[ADD_LDIST1]] = add nuw nsw i64 [[IND_LDIST1]], 1186; CHECK-NEXT: [[INC1_LDIST1]] = add i32 [[IND1_LDIST1]], 1187; CHECK-NEXT: [[ARRAYIDXA_PLUS_4_LDIST1:%.*]] = getelementptr inbounds i32, ptr [[A]], i64 [[ADD_LDIST1]]188; CHECK-NEXT: store i32 [[MULA_LDIST1]], ptr [[ARRAYIDXA_PLUS_4_LDIST1]], align 4, !alias.scope [[META11:![0-9]+]]189; CHECK-NEXT: [[EXITCOND_LDIST1:%.*]] = icmp eq i64 [[ADD_LDIST1]], [[N]]190; CHECK-NEXT: br i1 [[EXITCOND_LDIST1]], label [[FOR_BODY_PH:%.*]], label [[FOR_BODY_LDIST1]]191; CHECK: for.body.ph:192; CHECK-NEXT: br label [[FOR_BODY:%.*]]193; CHECK: for.body:194; CHECK-NEXT: [[IND:%.*]] = phi i64 [ 0, [[FOR_BODY_PH]] ], [ [[ADD:%.*]], [[FOR_BODY]] ]195; CHECK-NEXT: [[IND1:%.*]] = phi i32 [ 0, [[FOR_BODY_PH]] ], [ [[INC1:%.*]], [[FOR_BODY]] ]196; CHECK-NEXT: [[MUL:%.*]] = mul i32 [[IND1]], 2197; CHECK-NEXT: [[MUL_EXT:%.*]] = zext i32 [[MUL]] to i64198; CHECK-NEXT: [[ADD]] = add nuw nsw i64 [[IND]], 1199; CHECK-NEXT: [[INC1]] = add i32 [[IND1]], 1200; CHECK-NEXT: [[ARRAYIDXD:%.*]] = getelementptr inbounds i32, ptr [[D]], i64 [[MUL_EXT]]201; CHECK-NEXT: [[LOADD:%.*]] = load i32, ptr [[ARRAYIDXD]], align 4202; CHECK-NEXT: [[ARRAYIDXE:%.*]] = getelementptr inbounds i32, ptr [[E]], i64 [[MUL_EXT]]203; CHECK-NEXT: [[LOADE:%.*]] = load i32, ptr [[ARRAYIDXE]], align 4204; CHECK-NEXT: [[MULC:%.*]] = mul i32 [[LOADD]], [[LOADE]]205; CHECK-NEXT: [[ARRAYIDXC:%.*]] = getelementptr inbounds i32, ptr [[C]], i64 [[MUL_EXT]]206; CHECK-NEXT: store i32 [[MULC]], ptr [[ARRAYIDXC]], align 4207; CHECK-NEXT: [[EXITCOND:%.*]] = icmp eq i64 [[ADD]], [[N]]208; CHECK-NEXT: br i1 [[EXITCOND]], label [[FOR_END_LOOPEXIT2:%.*]], label [[FOR_BODY]]209; CHECK: for.end.loopexit:210; CHECK-NEXT: br label [[FOR_END:%.*]]211; CHECK: for.end.loopexit2:212; CHECK-NEXT: br label [[FOR_END]]213; CHECK: for.end:214; CHECK-NEXT: ret void215;216entry:217 %a_intptr = ptrtoint ptr @global_a to i64218 call void @use64(i64 %a_intptr)219 %a = getelementptr i32, ptr @global_a, i32 42220 br label %for.body221 222for.body: ; preds = %for.body, %entry223 %ind = phi i64 [ 0, %entry ], [ %add, %for.body ]224 %ind1 = phi i32 [ 0, %entry ], [ %inc1, %for.body ]225 226 %mul = mul i32 %ind1, 2227 %mul_ext = zext i32 %mul to i64228 229 230 %arrayidxA = getelementptr inbounds i32, ptr %a, i64 %mul_ext231 %loadA = load i32, ptr %arrayidxA, align 4232 233 %arrayidxB = getelementptr inbounds i32, ptr %b, i64 %mul_ext234 %loadB = load i32, ptr %arrayidxB, align 4235 236 %mulA = mul i32 %loadB, %loadA237 238 %add = add nuw nsw i64 %ind, 1239 %inc1 = add i32 %ind1, 1240 241 %arrayidxA_plus_4 = getelementptr inbounds i32, ptr %a, i64 %add242 store i32 %mulA, ptr %arrayidxA_plus_4, align 4243 244 %arrayidxD = getelementptr inbounds i32, ptr %d, i64 %mul_ext245 %loadD = load i32, ptr %arrayidxD, align 4246 247 %arrayidxE = getelementptr inbounds i32, ptr %e, i64 %mul_ext248 %loadE = load i32, ptr %arrayidxE, align 4249 250 %mulC = mul i32 %loadD, %loadE251 252 %arrayidxC = getelementptr inbounds i32, ptr %c, i64 %mul_ext253 store i32 %mulC, ptr %arrayidxC, align 4254 255 %exitcond = icmp eq i64 %add, %N256 br i1 %exitcond, label %for.end, label %for.body257 258for.end: ; preds = %for.body259 ret void260}261 262; Can't add control dependency with convergent in loop body.263define void @f_with_convergent(ptr noalias %a, ptr noalias %b, ptr noalias %c, ptr noalias %d, ptr noalias %e, i64 %N) #1 {264; CHECK-LABEL: @f_with_convergent(265; CHECK-NEXT: entry:266; CHECK-NEXT: br label [[FOR_BODY:%.*]]267; CHECK: for.body:268; CHECK-NEXT: [[IND:%.*]] = phi i64 [ 0, [[ENTRY:%.*]] ], [ [[ADD:%.*]], [[FOR_BODY]] ]269; CHECK-NEXT: [[IND1:%.*]] = phi i32 [ 0, [[ENTRY]] ], [ [[INC1:%.*]], [[FOR_BODY]] ]270; CHECK-NEXT: [[MUL:%.*]] = mul i32 [[IND1]], 2271; CHECK-NEXT: [[MUL_EXT:%.*]] = zext i32 [[MUL]] to i64272; CHECK-NEXT: [[ARRAYIDXA:%.*]] = getelementptr inbounds i32, ptr [[A:%.*]], i64 [[MUL_EXT]]273; CHECK-NEXT: [[LOADA:%.*]] = load i32, ptr [[ARRAYIDXA]], align 4274; CHECK-NEXT: [[ARRAYIDXB:%.*]] = getelementptr inbounds i32, ptr [[B:%.*]], i64 [[MUL_EXT]]275; CHECK-NEXT: [[LOADB:%.*]] = load i32, ptr [[ARRAYIDXB]], align 4276; CHECK-NEXT: [[MULA:%.*]] = mul i32 [[LOADB]], [[LOADA]]277; CHECK-NEXT: [[ADD]] = add nuw nsw i64 [[IND]], 1278; CHECK-NEXT: [[INC1]] = add i32 [[IND1]], 1279; CHECK-NEXT: [[ARRAYIDXA_PLUS_4:%.*]] = getelementptr inbounds i32, ptr [[A]], i64 [[ADD]]280; CHECK-NEXT: store i32 [[MULA]], ptr [[ARRAYIDXA_PLUS_4]], align 4281; CHECK-NEXT: [[ARRAYIDXD:%.*]] = getelementptr inbounds i32, ptr [[D:%.*]], i64 [[MUL_EXT]]282; CHECK-NEXT: [[LOADD:%.*]] = load i32, ptr [[ARRAYIDXD]], align 4283; CHECK-NEXT: [[ARRAYIDXE:%.*]] = getelementptr inbounds i32, ptr [[E:%.*]], i64 [[MUL_EXT]]284; CHECK-NEXT: [[LOADE:%.*]] = load i32, ptr [[ARRAYIDXE]], align 4285; CHECK-NEXT: [[CONVERGENTD:%.*]] = call i32 @llvm.convergent(i32 [[LOADD]])286; CHECK-NEXT: [[MULC:%.*]] = mul i32 [[CONVERGENTD]], [[LOADE]]287; CHECK-NEXT: [[ARRAYIDXC:%.*]] = getelementptr inbounds i32, ptr [[C:%.*]], i64 [[MUL_EXT]]288; CHECK-NEXT: store i32 [[MULC]], ptr [[ARRAYIDXC]], align 4289; CHECK-NEXT: [[EXITCOND:%.*]] = icmp eq i64 [[ADD]], [[N:%.*]]290; CHECK-NEXT: br i1 [[EXITCOND]], label [[FOR_END:%.*]], label [[FOR_BODY]]291; CHECK: for.end:292; CHECK-NEXT: ret void293;294entry:295 br label %for.body296 297for.body: ; preds = %for.body, %entry298 %ind = phi i64 [ 0, %entry ], [ %add, %for.body ]299 %ind1 = phi i32 [ 0, %entry ], [ %inc1, %for.body ]300 301 %mul = mul i32 %ind1, 2302 %mul_ext = zext i32 %mul to i64303 304 305 %arrayidxA = getelementptr inbounds i32, ptr %a, i64 %mul_ext306 %loadA = load i32, ptr %arrayidxA, align 4307 308 %arrayidxB = getelementptr inbounds i32, ptr %b, i64 %mul_ext309 %loadB = load i32, ptr %arrayidxB, align 4310 311 %mulA = mul i32 %loadB, %loadA312 313 %add = add nuw nsw i64 %ind, 1314 %inc1 = add i32 %ind1, 1315 316 %arrayidxA_plus_4 = getelementptr inbounds i32, ptr %a, i64 %add317 store i32 %mulA, ptr %arrayidxA_plus_4, align 4318 319 %arrayidxD = getelementptr inbounds i32, ptr %d, i64 %mul_ext320 %loadD = load i32, ptr %arrayidxD, align 4321 322 %arrayidxE = getelementptr inbounds i32, ptr %e, i64 %mul_ext323 %loadE = load i32, ptr %arrayidxE, align 4324 325 %convergentD = call i32 @llvm.convergent(i32 %loadD)326 %mulC = mul i32 %convergentD, %loadE327 328 %arrayidxC = getelementptr inbounds i32, ptr %c, i64 %mul_ext329 store i32 %mulC, ptr %arrayidxC, align 4330 331 %exitcond = icmp eq i64 %add, %N332 br i1 %exitcond, label %for.end, label %for.body333 334for.end: ; preds = %for.body335 ret void336}337 338declare i32 @llvm.convergent(i32) #0339 340attributes #0 = { nounwind readnone convergent }341attributes #1 = { nounwind convergent }342