236 lines · plain
1; NOTE: Assertions have been autogenerated by utils/update_test_checks.py UTC_ARGS: --version 32; RUN: opt -S -passes=loop-unroll -unroll-allow-partial < %s | FileCheck %s3 4target triple = "x86_64-unknown-linux-gnu"5 6define i32 @test(ptr %ary) "target-cpu"="znver3" {7; CHECK-LABEL: define i32 @test(8; CHECK-SAME: ptr [[ARY:%.*]]) #[[ATTR0:[0-9]+]] {9; CHECK-NEXT: entry:10; CHECK-NEXT: br label [[FOR_BODY:%.*]]11; CHECK: for.body:12; CHECK-NEXT: [[INDVARS_IV:%.*]] = phi i64 [ 0, [[ENTRY:%.*]] ], [ [[INDVARS_IV_NEXT_127:%.*]], [[FOR_BODY]] ]13; CHECK-NEXT: [[SUM:%.*]] = phi i32 [ 0, [[ENTRY]] ], [ [[SUM_NEXT_127:%.*]], [[FOR_BODY]] ]14; CHECK-NEXT: [[ARRAYIDX:%.*]] = getelementptr inbounds i32, ptr [[ARY]], i64 [[INDVARS_IV]]15; CHECK-NEXT: [[VAL:%.*]] = load i32, ptr [[ARRAYIDX]], align 416; CHECK-NEXT: [[SUM_NEXT:%.*]] = add nsw i32 [[VAL]], [[SUM]]17; CHECK-NEXT: [[INDVARS_IV_NEXT:%.*]] = add nuw nsw i64 [[INDVARS_IV]], 118; CHECK-NEXT: [[ARRAYIDX_1:%.*]] = getelementptr inbounds i32, ptr [[ARY]], i64 [[INDVARS_IV_NEXT]]19; CHECK-NEXT: [[VAL_1:%.*]] = load i32, ptr [[ARRAYIDX_1]], align 420; CHECK-NEXT: [[SUM_NEXT_1:%.*]] = add nsw i32 [[VAL_1]], [[SUM_NEXT]]21; CHECK-NEXT: [[INDVARS_IV_NEXT_1:%.*]] = add nuw nsw i64 [[INDVARS_IV]], 222; CHECK-NEXT: [[ARRAYIDX_2:%.*]] = getelementptr inbounds i32, ptr [[ARY]], i64 [[INDVARS_IV_NEXT_1]]23; CHECK-NEXT: [[VAL_2:%.*]] = load i32, ptr [[ARRAYIDX_2]], align 424; CHECK-NEXT: [[SUM_NEXT_2:%.*]] = add nsw i32 [[VAL_2]], [[SUM_NEXT_1]]25; CHECK-NEXT: [[INDVARS_IV_NEXT_2:%.*]] = add nuw nsw i64 [[INDVARS_IV]], 326; CHECK-NEXT: [[ARRAYIDX_3:%.*]] = getelementptr inbounds i32, ptr [[ARY]], i64 [[INDVARS_IV_NEXT_2]]27; CHECK-NEXT: [[VAL_3:%.*]] = load i32, ptr [[ARRAYIDX_3]], align 428; CHECK-NEXT: [[SUM_NEXT_3:%.*]] = add nsw i32 [[VAL_3]], [[SUM_NEXT_2]]29; CHECK-NEXT: [[INDVARS_IV_NEXT_3:%.*]] = add nuw nsw i64 [[INDVARS_IV]], 430; CHECK-NEXT: [[ARRAYIDX_4:%.*]] = getelementptr inbounds i32, ptr [[ARY]], i64 [[INDVARS_IV_NEXT_3]]31; CHECK-NEXT: [[VAL_4:%.*]] = load i32, ptr [[ARRAYIDX_4]], align 432; CHECK-NEXT: [[SUM_NEXT_4:%.*]] = add nsw i32 [[VAL_4]], [[SUM_NEXT_3]]33; CHECK-NEXT: [[INDVARS_IV_NEXT_4:%.*]] = add nuw nsw i64 [[INDVARS_IV]], 534; CHECK-NEXT: [[ARRAYIDX_5:%.*]] = getelementptr inbounds i32, ptr [[ARY]], i64 [[INDVARS_IV_NEXT_4]]35; CHECK-NEXT: [[VAL_5:%.*]] = load i32, ptr [[ARRAYIDX_5]], align 436; CHECK-NEXT: [[SUM_NEXT_5:%.*]] = add nsw i32 [[VAL_5]], [[SUM_NEXT_4]]37; CHECK-NEXT: [[INDVARS_IV_NEXT_5:%.*]] = add nuw nsw i64 [[INDVARS_IV]], 638; CHECK-NEXT: [[ARRAYIDX_6:%.*]] = getelementptr inbounds i32, ptr [[ARY]], i64 [[INDVARS_IV_NEXT_5]]39; CHECK-NEXT: [[VAL_6:%.*]] = load i32, ptr [[ARRAYIDX_6]], align 440; CHECK-NEXT: [[SUM_NEXT_6:%.*]] = add nsw i32 [[VAL_6]], [[SUM_NEXT_5]]41; CHECK-NEXT: [[INDVARS_IV_NEXT_6:%.*]] = add nuw nsw i64 [[INDVARS_IV]], 742; CHECK-NEXT: [[ARRAYIDX_7:%.*]] = getelementptr inbounds i32, ptr [[ARY]], i64 [[INDVARS_IV_NEXT_6]]43; CHECK-NEXT: [[VAL_7:%.*]] = load i32, ptr [[ARRAYIDX_7]], align 444; CHECK-NEXT: [[SUM_NEXT_7:%.*]] = add nsw i32 [[VAL_7]], [[SUM_NEXT_6]]45; CHECK-NEXT: [[INDVARS_IV_NEXT_7:%.*]] = add nuw nsw i64 [[INDVARS_IV]], 846; CHECK-NEXT: [[ARRAYIDX_8:%.*]] = getelementptr inbounds i32, ptr [[ARY]], i64 [[INDVARS_IV_NEXT_7]]47; CHECK-NEXT: [[VAL_8:%.*]] = load i32, ptr [[ARRAYIDX_8]], align 448; CHECK-NEXT: [[SUM_NEXT_8:%.*]] = add nsw i32 [[VAL_8]], [[SUM_NEXT_7]]49; CHECK-NEXT: [[INDVARS_IV_NEXT_8:%.*]] = add nuw nsw i64 [[INDVARS_IV]], 950; CHECK-NEXT: [[ARRAYIDX_9:%.*]] = getelementptr inbounds i32, ptr [[ARY]], i64 [[INDVARS_IV_NEXT_8]]51; CHECK-NEXT: [[VAL_9:%.*]] = load i32, ptr [[ARRAYIDX_9]], align 452; CHECK-NEXT: [[SUM_NEXT_9:%.*]] = add nsw i32 [[VAL_9]], [[SUM_NEXT_8]]53; CHECK-NEXT: [[INDVARS_IV_NEXT_9:%.*]] = add nuw nsw i64 [[INDVARS_IV]], 1054; CHECK-NEXT: [[ARRAYIDX_10:%.*]] = getelementptr inbounds i32, ptr [[ARY]], i64 [[INDVARS_IV_NEXT_9]]55; CHECK-NEXT: [[VAL_10:%.*]] = load i32, ptr [[ARRAYIDX_10]], align 456; CHECK-NEXT: [[SUM_NEXT_10:%.*]] = add nsw i32 [[VAL_10]], [[SUM_NEXT_9]]57; CHECK-NEXT: [[INDVARS_IV_NEXT_10:%.*]] = add nuw nsw i64 [[INDVARS_IV]], 1158; CHECK-NEXT: [[ARRAYIDX_11:%.*]] = getelementptr inbounds i32, ptr [[ARY]], i64 [[INDVARS_IV_NEXT_10]]59; CHECK-NEXT: [[VAL_11:%.*]] = load i32, ptr [[ARRAYIDX_11]], align 460; CHECK-NEXT: [[SUM_NEXT_11:%.*]] = add nsw i32 [[VAL_11]], [[SUM_NEXT_10]]61; CHECK-NEXT: [[INDVARS_IV_NEXT_11:%.*]] = add nuw nsw i64 [[INDVARS_IV]], 1262; CHECK-NEXT: [[ARRAYIDX_12:%.*]] = getelementptr inbounds i32, ptr [[ARY]], i64 [[INDVARS_IV_NEXT_11]]63; CHECK-NEXT: [[VAL_12:%.*]] = load i32, ptr [[ARRAYIDX_12]], align 464; CHECK-NEXT: [[SUM_NEXT_12:%.*]] = add nsw i32 [[VAL_12]], [[SUM_NEXT_11]]65; CHECK-NEXT: [[INDVARS_IV_NEXT_12:%.*]] = add nuw nsw i64 [[INDVARS_IV]], 1366; CHECK-NEXT: [[ARRAYIDX_13:%.*]] = getelementptr inbounds i32, ptr [[ARY]], i64 [[INDVARS_IV_NEXT_12]]67; CHECK-NEXT: [[VAL_13:%.*]] = load i32, ptr [[ARRAYIDX_13]], align 468; CHECK-NEXT: [[SUM_NEXT_13:%.*]] = add nsw i32 [[VAL_13]], [[SUM_NEXT_12]]69; CHECK-NEXT: [[INDVARS_IV_NEXT_13:%.*]] = add nuw nsw i64 [[INDVARS_IV]], 1470; CHECK-NEXT: [[ARRAYIDX_14:%.*]] = getelementptr inbounds i32, ptr [[ARY]], i64 [[INDVARS_IV_NEXT_13]]71; CHECK-NEXT: [[VAL_14:%.*]] = load i32, ptr [[ARRAYIDX_14]], align 472; CHECK-NEXT: [[SUM_NEXT_14:%.*]] = add nsw i32 [[VAL_14]], [[SUM_NEXT_13]]73; CHECK-NEXT: [[INDVARS_IV_NEXT_14:%.*]] = add nuw nsw i64 [[INDVARS_IV]], 1574; CHECK-NEXT: [[ARRAYIDX_15:%.*]] = getelementptr inbounds i32, ptr [[ARY]], i64 [[INDVARS_IV_NEXT_14]]75; CHECK-NEXT: [[VAL_15:%.*]] = load i32, ptr [[ARRAYIDX_15]], align 476; CHECK-NEXT: [[SUM_NEXT_127]] = add nsw i32 [[VAL_15]], [[SUM_NEXT_14]]77; CHECK-NEXT: [[INDVARS_IV_NEXT_127]] = add nuw nsw i64 [[INDVARS_IV]], 1678; CHECK-NEXT: [[EXITCOND_NOT_127:%.*]] = icmp eq i64 [[INDVARS_IV_NEXT_127]], 819279; CHECK-NEXT: br i1 [[EXITCOND_NOT_127]], label [[FOR_COND_CLEANUP:%.*]], label [[FOR_BODY]]80; CHECK: for.cond.cleanup:81; CHECK-NEXT: [[SUM_NEXT_LCSSA:%.*]] = phi i32 [ [[SUM_NEXT_127]], [[FOR_BODY]] ]82; CHECK-NEXT: ret i32 [[SUM_NEXT_LCSSA]]83;84entry:85 br label %for.body86 87for.body:88 %indvars.iv = phi i64 [ 0, %entry ], [ %indvars.iv.next, %for.body ]89 %sum = phi i32 [ 0, %entry ], [ %sum.next, %for.body ]90 %arrayidx = getelementptr inbounds i32, ptr %ary, i64 %indvars.iv91 %val = load i32, ptr %arrayidx, align 492 %sum.next = add nsw i32 %val, %sum93 %indvars.iv.next = add nuw nsw i64 %indvars.iv, 194 %exitcond.not = icmp eq i64 %indvars.iv.next, 819295 br i1 %exitcond.not, label %for.cond.cleanup, label %for.body96 97for.cond.cleanup:98 ret i32 %sum.next99}100 101define i32 @test2(ptr %ary, i64 %n) "target-cpu"="znver3" {102; CHECK-LABEL: define i32 @test2(103; CHECK-SAME: ptr [[ARY:%.*]], i64 [[N:%.*]]) #[[ATTR0]] {104; CHECK-NEXT: entry:105; CHECK-NEXT: br label [[FOR_BODY:%.*]]106; CHECK: for.body:107; CHECK-NEXT: [[INDVARS_IV_EPIL:%.*]] = phi i64 [ 0, [[ENTRY:%.*]] ], [ [[INDVARS_IV_NEXT_EPIL:%.*]], [[FOR_BODY]] ]108; CHECK-NEXT: [[SUM_EPIL:%.*]] = phi i32 [ 0, [[ENTRY]] ], [ [[SUM_NEXT_EPIL:%.*]], [[FOR_BODY]] ]109; CHECK-NEXT: [[ARRAYIDX_EPIL:%.*]] = getelementptr inbounds i32, ptr [[ARY]], i64 [[INDVARS_IV_EPIL]]110; CHECK-NEXT: [[VAL_EPIL:%.*]] = load i32, ptr [[ARRAYIDX_EPIL]], align 4111; CHECK-NEXT: [[DUMMY21:%.*]] = mul i32 [[VAL_EPIL]], [[VAL_EPIL]]112; CHECK-NEXT: [[DUMMY27:%.*]] = mul i32 [[DUMMY21]], [[DUMMY21]]113; CHECK-NEXT: [[DUMMY28:%.*]] = mul i32 [[DUMMY27]], [[DUMMY27]]114; CHECK-NEXT: [[DUMMY4:%.*]] = mul i32 [[DUMMY28]], [[DUMMY28]]115; CHECK-NEXT: [[DUMMY5:%.*]] = mul i32 [[DUMMY4]], [[DUMMY4]]116; CHECK-NEXT: [[DUMMY6:%.*]] = mul i32 [[DUMMY5]], [[DUMMY5]]117; CHECK-NEXT: [[DUMMY7:%.*]] = mul i32 [[DUMMY6]], [[DUMMY6]]118; CHECK-NEXT: [[DUMMY8:%.*]] = mul i32 [[DUMMY7]], [[DUMMY7]]119; CHECK-NEXT: [[DUMMY9:%.*]] = mul i32 [[DUMMY8]], [[DUMMY8]]120; CHECK-NEXT: [[DUMMY10:%.*]] = mul i32 [[DUMMY9]], [[DUMMY9]]121; CHECK-NEXT: [[DUMMY29:%.*]] = mul i32 [[DUMMY10]], [[DUMMY10]]122; CHECK-NEXT: [[DUMMY30:%.*]] = mul i32 [[DUMMY29]], [[DUMMY29]]123; CHECK-NEXT: [[DUMMY39:%.*]] = mul i32 [[DUMMY30]], [[DUMMY30]]124; CHECK-NEXT: [[DUMMY40:%.*]] = mul i32 [[DUMMY39]], [[DUMMY39]]125; CHECK-NEXT: [[DUMMY15:%.*]] = mul i32 [[DUMMY40]], [[DUMMY40]]126; CHECK-NEXT: [[DUMMY16:%.*]] = mul i32 [[DUMMY15]], [[DUMMY15]]127; CHECK-NEXT: [[DUMMY17:%.*]] = mul i32 [[DUMMY16]], [[DUMMY16]]128; CHECK-NEXT: [[DUMMY18:%.*]] = mul i32 [[DUMMY17]], [[DUMMY17]]129; CHECK-NEXT: [[DUMMY19:%.*]] = mul i32 [[DUMMY18]], [[DUMMY18]]130; CHECK-NEXT: [[DUMMY20:%.*]] = mul i32 [[DUMMY19]], [[DUMMY19]]131; CHECK-NEXT: [[VAL:%.*]] = mul i32 [[DUMMY20]], [[DUMMY20]]132; CHECK-NEXT: [[DUMMY1:%.*]] = mul i32 [[VAL]], [[VAL]]133; CHECK-NEXT: [[DUMMY2:%.*]] = mul i32 [[DUMMY1]], [[DUMMY1]]134; CHECK-NEXT: [[DUMMY3:%.*]] = mul i32 [[DUMMY2]], [[DUMMY2]]135; CHECK-NEXT: [[DUMMY41:%.*]] = mul i32 [[DUMMY3]], [[DUMMY3]]136; CHECK-NEXT: [[DUMMY26:%.*]] = mul i32 [[DUMMY41]], [[DUMMY41]]137; CHECK-NEXT: [[DUMMY11:%.*]] = mul i32 [[DUMMY26]], [[DUMMY26]]138; CHECK-NEXT: [[DUMMY12:%.*]] = mul i32 [[DUMMY11]], [[DUMMY11]]139; CHECK-NEXT: [[DUMMY13:%.*]] = mul i32 [[DUMMY12]], [[DUMMY12]]140; CHECK-NEXT: [[DUMMY14:%.*]] = mul i32 [[DUMMY13]], [[DUMMY13]]141; CHECK-NEXT: [[DUMMY31:%.*]] = mul i32 [[DUMMY14]], [[DUMMY14]]142; CHECK-NEXT: [[DUMMY32:%.*]] = mul i32 [[DUMMY31]], [[DUMMY31]]143; CHECK-NEXT: [[DUMMY22:%.*]] = mul i32 [[DUMMY32]], [[DUMMY32]]144; CHECK-NEXT: [[DUMMY23:%.*]] = mul i32 [[DUMMY22]], [[DUMMY22]]145; CHECK-NEXT: [[DUMMY24:%.*]] = mul i32 [[DUMMY23]], [[DUMMY23]]146; CHECK-NEXT: [[DUMMY25:%.*]] = mul i32 [[DUMMY24]], [[DUMMY24]]147; CHECK-NEXT: [[DUMMY37:%.*]] = mul i32 [[DUMMY25]], [[DUMMY25]]148; CHECK-NEXT: [[DUMMY38:%.*]] = mul i32 [[DUMMY37]], [[DUMMY37]]149; CHECK-NEXT: [[DUMMY33:%.*]] = mul i32 [[DUMMY38]], [[DUMMY38]]150; CHECK-NEXT: [[DUMMY34:%.*]] = mul i32 [[DUMMY33]], [[DUMMY33]]151; CHECK-NEXT: [[DUMMY35:%.*]] = mul i32 [[DUMMY34]], [[DUMMY34]]152; CHECK-NEXT: [[DUMMY36:%.*]] = mul i32 [[DUMMY35]], [[DUMMY35]]153; CHECK-NEXT: [[DUMMY43:%.*]] = mul i32 [[DUMMY36]], [[DUMMY36]]154; CHECK-NEXT: [[DUMMY44:%.*]] = mul i32 [[DUMMY43]], [[DUMMY43]]155; CHECK-NEXT: [[DUMMY45:%.*]] = mul i32 [[DUMMY44]], [[DUMMY44]]156; CHECK-NEXT: [[DUMMY46:%.*]] = mul i32 [[DUMMY45]], [[DUMMY45]]157; CHECK-NEXT: [[DUMMY47:%.*]] = mul i32 [[DUMMY46]], [[DUMMY46]]158; CHECK-NEXT: [[DUMMY48:%.*]] = mul i32 [[DUMMY47]], [[DUMMY47]]159; CHECK-NEXT: [[DUMMY49:%.*]] = mul i32 [[DUMMY48]], [[DUMMY48]]160; CHECK-NEXT: [[DUMMY50_EPIL:%.*]] = mul i32 [[DUMMY49]], [[DUMMY49]]161; CHECK-NEXT: [[SUM_NEXT_EPIL]] = add nsw i32 [[DUMMY50_EPIL]], [[SUM_EPIL]]162; CHECK-NEXT: [[INDVARS_IV_NEXT_EPIL]] = add nuw nsw i64 [[INDVARS_IV_EPIL]], 1163; CHECK-NEXT: [[EXITCOND_NOT_EPIL:%.*]] = icmp eq i64 [[INDVARS_IV_NEXT_EPIL]], [[N]]164; CHECK-NEXT: br i1 [[EXITCOND_NOT_EPIL]], label [[FOR_COND_CLEANUP:%.*]], label [[FOR_BODY]]165; CHECK: for.cond.cleanup:166; CHECK-NEXT: [[SUM_NEXT_LCSSA:%.*]] = phi i32 [ [[SUM_NEXT_EPIL]], [[FOR_BODY]] ]167; CHECK-NEXT: ret i32 [[SUM_NEXT_LCSSA]]168;169entry:170 br label %for.body171 172for.body:173 %indvars.iv = phi i64 [ 0, %entry ], [ %indvars.iv.next, %for.body ]174 %sum = phi i32 [ 0, %entry ], [ %sum.next, %for.body ]175 %arrayidx = getelementptr inbounds i32, ptr %ary, i64 %indvars.iv176 %val = load i32, ptr %arrayidx, align 4177 %dummy1 = mul i32 %val, %val178 %dummy2 = mul i32 %dummy1, %dummy1179 %dummy3 = mul i32 %dummy2, %dummy2180 %dummy4 = mul i32 %dummy3, %dummy3181 %dummy5 = mul i32 %dummy4, %dummy4182 %dummy6 = mul i32 %dummy5, %dummy5183 %dummy7 = mul i32 %dummy6, %dummy6184 %dummy8 = mul i32 %dummy7, %dummy7185 %dummy9 = mul i32 %dummy8, %dummy8186 %dummy10 = mul i32 %dummy9, %dummy9187 %dummy11 = mul i32 %dummy10, %dummy10188 %dummy12 = mul i32 %dummy11, %dummy11189 %dummy13 = mul i32 %dummy12, %dummy12190 %dummy14 = mul i32 %dummy13, %dummy13191 %dummy15 = mul i32 %dummy14, %dummy14192 %dummy16 = mul i32 %dummy15, %dummy15193 %dummy17 = mul i32 %dummy16, %dummy16194 %dummy18 = mul i32 %dummy17, %dummy17195 %dummy19 = mul i32 %dummy18, %dummy18196 %dummy20 = mul i32 %dummy19, %dummy19197 %dummy21 = mul i32 %dummy20, %dummy20198 %dummy22 = mul i32 %dummy21, %dummy21199 %dummy23 = mul i32 %dummy22, %dummy22200 %dummy24 = mul i32 %dummy23, %dummy23201 %dummy25 = mul i32 %dummy24, %dummy24202 %dummy26 = mul i32 %dummy25, %dummy25203 %dummy27 = mul i32 %dummy26, %dummy26204 %dummy28 = mul i32 %dummy27, %dummy27205 %dummy29 = mul i32 %dummy28, %dummy28206 %dummy30 = mul i32 %dummy29, %dummy29207 %dummy31 = mul i32 %dummy30, %dummy30208 %dummy32 = mul i32 %dummy31, %dummy31209 %dummy33 = mul i32 %dummy32, %dummy32210 %dummy34 = mul i32 %dummy33, %dummy33211 %dummy35 = mul i32 %dummy34, %dummy34212 %dummy36 = mul i32 %dummy35, %dummy35213 %dummy37 = mul i32 %dummy36, %dummy36214 %dummy38 = mul i32 %dummy37, %dummy37215 %dummy39 = mul i32 %dummy38, %dummy38216 %dummy40 = mul i32 %dummy39, %dummy39217 %dummy41 = mul i32 %dummy40, %dummy40218 %dummy42 = mul i32 %dummy41, %dummy41219 %dummy43 = mul i32 %dummy42, %dummy42220 %dummy44 = mul i32 %dummy43, %dummy43221 %dummy45 = mul i32 %dummy44, %dummy44222 %dummy46 = mul i32 %dummy45, %dummy45223 %dummy47 = mul i32 %dummy46, %dummy46224 %dummy48 = mul i32 %dummy47, %dummy47225 %dummy49 = mul i32 %dummy48, %dummy48226 %dummy50 = mul i32 %dummy49, %dummy49227 %sum.next = add nsw i32 %dummy50, %sum228 %indvars.iv.next = add nuw nsw i64 %indvars.iv, 1229 %exitcond.not = icmp eq i64 %indvars.iv.next, %n230 br i1 %exitcond.not, label %for.cond.cleanup, label %for.body231 232for.cond.cleanup:233 ret i32 %sum.next234}235 236