146 lines · plain
1; NOTE: Assertions have been autogenerated by utils/update_test_checks.py2; RUN: opt -passes=loop-distribute -enable-loop-distribute -S < %s | \3; RUN: FileCheck %s --check-prefix=DEFAULT4 5; RUN: opt -passes=loop-distribute -enable-loop-distribute -S -enable-mem-access-versioning=0 < %s | \6; RUN: FileCheck %s --check-prefix=NO-VERSION7 8; If we don't speculate stride for 1 we can't distribute along the line9; because we could have a backward dependence:10;11; for (i = 0; i < n; i++) {12; A[i + 1] = A[i] * B[i];13; =======================14; C[i] = D[i] * A[stride * i];15; }16 17define void @f(ptr noalias %a, ptr noalias %b, ptr noalias %c, ptr noalias %d, i64 %stride) {18; DEFAULT-LABEL: @f(19; DEFAULT-NEXT: entry:20; DEFAULT-NEXT: br label [[FOR_BODY_LVER_CHECK:%.*]]21; DEFAULT: for.body.lver.check:22; DEFAULT-NEXT: [[IDENT_CHECK:%.*]] = icmp ne i64 [[STRIDE:%.*]], 123; DEFAULT-NEXT: br i1 [[IDENT_CHECK]], label [[FOR_BODY_PH_LVER_ORIG:%.*]], label [[FOR_BODY_PH_LDIST1:%.*]]24; DEFAULT: for.body.ph.lver.orig:25; DEFAULT-NEXT: br label [[FOR_BODY_LVER_ORIG:%.*]]26; DEFAULT: for.body.lver.orig:27; DEFAULT-NEXT: [[IND_LVER_ORIG:%.*]] = phi i64 [ 0, [[FOR_BODY_PH_LVER_ORIG]] ], [ [[ADD_LVER_ORIG:%.*]], [[FOR_BODY_LVER_ORIG]] ]28; DEFAULT-NEXT: [[ARRAYIDXA_LVER_ORIG:%.*]] = getelementptr inbounds i32, ptr [[A:%.*]], i64 [[IND_LVER_ORIG]]29; DEFAULT-NEXT: [[LOADA_LVER_ORIG:%.*]] = load i32, ptr [[ARRAYIDXA_LVER_ORIG]], align 430; DEFAULT-NEXT: [[ARRAYIDXB_LVER_ORIG:%.*]] = getelementptr inbounds i32, ptr [[B:%.*]], i64 [[IND_LVER_ORIG]]31; DEFAULT-NEXT: [[LOADB_LVER_ORIG:%.*]] = load i32, ptr [[ARRAYIDXB_LVER_ORIG]], align 432; DEFAULT-NEXT: [[MULA_LVER_ORIG:%.*]] = mul i32 [[LOADB_LVER_ORIG]], [[LOADA_LVER_ORIG]]33; DEFAULT-NEXT: [[ADD_LVER_ORIG]] = add nuw nsw i64 [[IND_LVER_ORIG]], 134; DEFAULT-NEXT: [[ARRAYIDXA_PLUS_4_LVER_ORIG:%.*]] = getelementptr inbounds i32, ptr [[A]], i64 [[ADD_LVER_ORIG]]35; DEFAULT-NEXT: store i32 [[MULA_LVER_ORIG]], ptr [[ARRAYIDXA_PLUS_4_LVER_ORIG]], align 436; DEFAULT-NEXT: [[ARRAYIDXD_LVER_ORIG:%.*]] = getelementptr inbounds i32, ptr [[D:%.*]], i64 [[IND_LVER_ORIG]]37; DEFAULT-NEXT: [[LOADD_LVER_ORIG:%.*]] = load i32, ptr [[ARRAYIDXD_LVER_ORIG]], align 438; DEFAULT-NEXT: [[MUL_LVER_ORIG:%.*]] = mul i64 [[IND_LVER_ORIG]], [[STRIDE]]39; DEFAULT-NEXT: [[ARRAYIDXSTRIDEDA_LVER_ORIG:%.*]] = getelementptr inbounds i32, ptr [[A]], i64 [[MUL_LVER_ORIG]]40; DEFAULT-NEXT: [[LOADSTRIDEDA_LVER_ORIG:%.*]] = load i32, ptr [[ARRAYIDXSTRIDEDA_LVER_ORIG]], align 441; DEFAULT-NEXT: [[MULC_LVER_ORIG:%.*]] = mul i32 [[LOADD_LVER_ORIG]], [[LOADSTRIDEDA_LVER_ORIG]]42; DEFAULT-NEXT: [[ARRAYIDXC_LVER_ORIG:%.*]] = getelementptr inbounds i32, ptr [[C:%.*]], i64 [[IND_LVER_ORIG]]43; DEFAULT-NEXT: store i32 [[MULC_LVER_ORIG]], ptr [[ARRAYIDXC_LVER_ORIG]], align 444; DEFAULT-NEXT: [[EXITCOND_LVER_ORIG:%.*]] = icmp eq i64 [[ADD_LVER_ORIG]], 2045; DEFAULT-NEXT: br i1 [[EXITCOND_LVER_ORIG]], label [[FOR_END_LOOPEXIT:%.*]], label [[FOR_BODY_LVER_ORIG]]46; DEFAULT: for.body.ph.ldist1:47; DEFAULT-NEXT: br label [[FOR_BODY_LDIST1:%.*]]48; DEFAULT: for.body.ldist1:49; DEFAULT-NEXT: [[IND_LDIST1:%.*]] = phi i64 [ 0, [[FOR_BODY_PH_LDIST1]] ], [ [[ADD_LDIST1:%.*]], [[FOR_BODY_LDIST1]] ]50; DEFAULT-NEXT: [[ARRAYIDXA_LDIST1:%.*]] = getelementptr inbounds i32, ptr [[A]], i64 [[IND_LDIST1]]51; DEFAULT-NEXT: [[LOADA_LDIST1:%.*]] = load i32, ptr [[ARRAYIDXA_LDIST1]], align 452; DEFAULT-NEXT: [[ARRAYIDXB_LDIST1:%.*]] = getelementptr inbounds i32, ptr [[B]], i64 [[IND_LDIST1]]53; DEFAULT-NEXT: [[LOADB_LDIST1:%.*]] = load i32, ptr [[ARRAYIDXB_LDIST1]], align 454; DEFAULT-NEXT: [[MULA_LDIST1:%.*]] = mul i32 [[LOADB_LDIST1]], [[LOADA_LDIST1]]55; DEFAULT-NEXT: [[ADD_LDIST1]] = add nuw nsw i64 [[IND_LDIST1]], 156; DEFAULT-NEXT: [[ARRAYIDXA_PLUS_4_LDIST1:%.*]] = getelementptr inbounds i32, ptr [[A]], i64 [[ADD_LDIST1]]57; DEFAULT-NEXT: store i32 [[MULA_LDIST1]], ptr [[ARRAYIDXA_PLUS_4_LDIST1]], align 458; DEFAULT-NEXT: [[EXITCOND_LDIST1:%.*]] = icmp eq i64 [[ADD_LDIST1]], 2059; DEFAULT-NEXT: br i1 [[EXITCOND_LDIST1]], label [[FOR_BODY_PH:%.*]], label [[FOR_BODY_LDIST1]]60; DEFAULT: for.body.ph:61; DEFAULT-NEXT: br label [[FOR_BODY:%.*]]62; DEFAULT: for.body:63; DEFAULT-NEXT: [[IND:%.*]] = phi i64 [ 0, [[FOR_BODY_PH]] ], [ [[ADD:%.*]], [[FOR_BODY]] ]64; DEFAULT-NEXT: [[ADD]] = add nuw nsw i64 [[IND]], 165; DEFAULT-NEXT: [[ARRAYIDXD:%.*]] = getelementptr inbounds i32, ptr [[D]], i64 [[IND]]66; DEFAULT-NEXT: [[LOADD:%.*]] = load i32, ptr [[ARRAYIDXD]], align 467; DEFAULT-NEXT: [[MUL:%.*]] = mul i64 [[IND]], [[STRIDE]]68; DEFAULT-NEXT: [[ARRAYIDXSTRIDEDA:%.*]] = getelementptr inbounds i32, ptr [[A]], i64 [[MUL]]69; DEFAULT-NEXT: [[LOADSTRIDEDA:%.*]] = load i32, ptr [[ARRAYIDXSTRIDEDA]], align 470; DEFAULT-NEXT: [[MULC:%.*]] = mul i32 [[LOADD]], [[LOADSTRIDEDA]]71; DEFAULT-NEXT: [[ARRAYIDXC:%.*]] = getelementptr inbounds i32, ptr [[C]], i64 [[IND]]72; DEFAULT-NEXT: store i32 [[MULC]], ptr [[ARRAYIDXC]], align 473; DEFAULT-NEXT: [[EXITCOND:%.*]] = icmp eq i64 [[ADD]], 2074; DEFAULT-NEXT: br i1 [[EXITCOND]], label [[FOR_END_LOOPEXIT1:%.*]], label [[FOR_BODY]]75; DEFAULT: for.end.loopexit:76; DEFAULT-NEXT: br label [[FOR_END:%.*]]77; DEFAULT: for.end.loopexit1:78; DEFAULT-NEXT: br label [[FOR_END]]79; DEFAULT: for.end:80; DEFAULT-NEXT: ret void81;82; NO-VERSION-LABEL: @f(83; NO-VERSION-NEXT: entry:84; NO-VERSION-NEXT: br label [[FOR_BODY:%.*]]85; NO-VERSION: for.body:86; NO-VERSION-NEXT: [[IND:%.*]] = phi i64 [ 0, [[ENTRY:%.*]] ], [ [[ADD:%.*]], [[FOR_BODY]] ]87; NO-VERSION-NEXT: [[ARRAYIDXA:%.*]] = getelementptr inbounds i32, ptr [[A:%.*]], i64 [[IND]]88; NO-VERSION-NEXT: [[LOADA:%.*]] = load i32, ptr [[ARRAYIDXA]], align 489; NO-VERSION-NEXT: [[ARRAYIDXB:%.*]] = getelementptr inbounds i32, ptr [[B:%.*]], i64 [[IND]]90; NO-VERSION-NEXT: [[LOADB:%.*]] = load i32, ptr [[ARRAYIDXB]], align 491; NO-VERSION-NEXT: [[MULA:%.*]] = mul i32 [[LOADB]], [[LOADA]]92; NO-VERSION-NEXT: [[ADD]] = add nuw nsw i64 [[IND]], 193; NO-VERSION-NEXT: [[ARRAYIDXA_PLUS_4:%.*]] = getelementptr inbounds i32, ptr [[A]], i64 [[ADD]]94; NO-VERSION-NEXT: store i32 [[MULA]], ptr [[ARRAYIDXA_PLUS_4]], align 495; NO-VERSION-NEXT: [[ARRAYIDXD:%.*]] = getelementptr inbounds i32, ptr [[D:%.*]], i64 [[IND]]96; NO-VERSION-NEXT: [[LOADD:%.*]] = load i32, ptr [[ARRAYIDXD]], align 497; NO-VERSION-NEXT: [[MUL:%.*]] = mul i64 [[IND]], [[STRIDE:%.*]]98; NO-VERSION-NEXT: [[ARRAYIDXSTRIDEDA:%.*]] = getelementptr inbounds i32, ptr [[A]], i64 [[MUL]]99; NO-VERSION-NEXT: [[LOADSTRIDEDA:%.*]] = load i32, ptr [[ARRAYIDXSTRIDEDA]], align 4100; NO-VERSION-NEXT: [[MULC:%.*]] = mul i32 [[LOADD]], [[LOADSTRIDEDA]]101; NO-VERSION-NEXT: [[ARRAYIDXC:%.*]] = getelementptr inbounds i32, ptr [[C:%.*]], i64 [[IND]]102; NO-VERSION-NEXT: store i32 [[MULC]], ptr [[ARRAYIDXC]], align 4103; NO-VERSION-NEXT: [[EXITCOND:%.*]] = icmp eq i64 [[ADD]], 20104; NO-VERSION-NEXT: br i1 [[EXITCOND]], label [[FOR_END:%.*]], label [[FOR_BODY]]105; NO-VERSION: for.end:106; NO-VERSION-NEXT: ret void107;108entry:109 br label %for.body110 111 112 113for.body: ; preds = %for.body, %entry114 %ind = phi i64 [ 0, %entry ], [ %add, %for.body ]115 116 %arrayidxA = getelementptr inbounds i32, ptr %a, i64 %ind117 %loadA = load i32, ptr %arrayidxA, align 4118 119 %arrayidxB = getelementptr inbounds i32, ptr %b, i64 %ind120 %loadB = load i32, ptr %arrayidxB, align 4121 122 %mulA = mul i32 %loadB, %loadA123 124 %add = add nuw nsw i64 %ind, 1125 %arrayidxA_plus_4 = getelementptr inbounds i32, ptr %a, i64 %add126 store i32 %mulA, ptr %arrayidxA_plus_4, align 4127 128 %arrayidxD = getelementptr inbounds i32, ptr %d, i64 %ind129 %loadD = load i32, ptr %arrayidxD, align 4130 131 %mul = mul i64 %ind, %stride132 %arrayidxStridedA = getelementptr inbounds i32, ptr %a, i64 %mul133 %loadStridedA = load i32, ptr %arrayidxStridedA, align 4134 135 %mulC = mul i32 %loadD, %loadStridedA136 137 %arrayidxC = getelementptr inbounds i32, ptr %c, i64 %ind138 store i32 %mulC, ptr %arrayidxC, align 4139 140 %exitcond = icmp eq i64 %add, 20141 br i1 %exitcond, label %for.end, label %for.body142 143for.end: ; preds = %for.body144 ret void145}146