89 lines · plain
1; NOTE: Assertions have been autogenerated by utils/update_test_checks.py2; RUN: opt < %s -passes=slp-vectorizer -S -mtriple=i386-apple-macosx10.8.0 -mcpu=corei7-avx | FileCheck %s3 4define i32 @foo(ptr nocapture readonly %diff) #0 {5; CHECK-LABEL: @foo(6; CHECK-NEXT: entry:7; CHECK-NEXT: [[M2:%.*]] = alloca [8 x [8 x i32]], align 168; CHECK-NEXT: br label [[FOR_BODY:%.*]]9; CHECK: for.body:10; CHECK-NEXT: [[INDVARS_IV:%.*]] = phi i64 [ 0, [[ENTRY:%.*]] ], [ [[INDVARS_IV_NEXT:%.*]], [[FOR_BODY]] ]11; CHECK-NEXT: [[A_088:%.*]] = phi i32 [ 0, [[ENTRY]] ], [ [[OP_RDX:%.*]], [[FOR_BODY]] ]12; CHECK-NEXT: [[TMP0:%.*]] = shl i64 [[INDVARS_IV]], 313; CHECK-NEXT: [[ARRAYIDX:%.*]] = getelementptr inbounds i32, ptr [[DIFF:%.*]], i64 [[TMP0]]14; CHECK-NEXT: [[TMP1:%.*]] = or disjoint i64 [[TMP0]], 415; CHECK-NEXT: [[ARRAYIDX2:%.*]] = getelementptr inbounds i32, ptr [[DIFF]], i64 [[TMP1]]16; CHECK-NEXT: [[ARRAYIDX6:%.*]] = getelementptr inbounds [8 x [8 x i32]], ptr [[M2]], i64 0, i64 [[INDVARS_IV]], i64 017; CHECK-NEXT: [[TMP2:%.*]] = load <4 x i32>, ptr [[ARRAYIDX]], align 418; CHECK-NEXT: [[TMP3:%.*]] = load <4 x i32>, ptr [[ARRAYIDX2]], align 419; CHECK-NEXT: [[TMP4:%.*]] = add nsw <4 x i32> [[TMP3]], [[TMP2]]20; CHECK-NEXT: store <4 x i32> [[TMP4]], ptr [[ARRAYIDX6]], align 1621; CHECK-NEXT: [[TMP5:%.*]] = call i32 @llvm.vector.reduce.add.v4i32(<4 x i32> [[TMP4]])22; CHECK-NEXT: [[OP_RDX]] = add i32 [[TMP5]], [[A_088]]23; CHECK-NEXT: [[INDVARS_IV_NEXT]] = add nuw nsw i64 [[INDVARS_IV]], 124; CHECK-NEXT: [[EXITCOND:%.*]] = icmp eq i64 [[INDVARS_IV_NEXT]], 825; CHECK-NEXT: br i1 [[EXITCOND]], label [[FOR_END:%.*]], label [[FOR_BODY]]26; CHECK: for.end:27; CHECK-NEXT: call void @ff(ptr [[M2]])28; CHECK-NEXT: ret i32 [[OP_RDX]]29;30entry:31 %m2 = alloca [8 x [8 x i32]], align 1632 br label %for.body33 34for.body: ; preds = %for.body, %entry35 %indvars.iv = phi i64 [ 0, %entry ], [ %indvars.iv.next, %for.body ]36 %a.088 = phi i32 [ 0, %entry ], [ %add52, %for.body ]37 %0 = shl i64 %indvars.iv, 338 %arrayidx = getelementptr inbounds i32, ptr %diff, i64 %039 %1 = load i32, ptr %arrayidx, align 440 %2 = or disjoint i64 %0, 441 %arrayidx2 = getelementptr inbounds i32, ptr %diff, i64 %242 %3 = load i32, ptr %arrayidx2, align 443 %add3 = add nsw i32 %3, %144 %arrayidx6 = getelementptr inbounds [8 x [8 x i32]], ptr %m2, i64 0, i64 %indvars.iv, i64 045 store i32 %add3, ptr %arrayidx6, align 1646 %add10 = add nsw i32 %add3, %a.08847 %4 = or disjoint i64 %0, 148 %arrayidx13 = getelementptr inbounds i32, ptr %diff, i64 %449 %5 = load i32, ptr %arrayidx13, align 450 %6 = or disjoint i64 %0, 551 %arrayidx16 = getelementptr inbounds i32, ptr %diff, i64 %652 %7 = load i32, ptr %arrayidx16, align 453 %add17 = add nsw i32 %7, %554 %arrayidx20 = getelementptr inbounds [8 x [8 x i32]], ptr %m2, i64 0, i64 %indvars.iv, i64 155 store i32 %add17, ptr %arrayidx20, align 456 %add24 = add nsw i32 %add10, %add1757 %8 = or disjoint i64 %0, 258 %arrayidx27 = getelementptr inbounds i32, ptr %diff, i64 %859 %9 = load i32, ptr %arrayidx27, align 460 %10 = or disjoint i64 %0, 661 %arrayidx30 = getelementptr inbounds i32, ptr %diff, i64 %1062 %11 = load i32, ptr %arrayidx30, align 463 %add31 = add nsw i32 %11, %964 %arrayidx34 = getelementptr inbounds [8 x [8 x i32]], ptr %m2, i64 0, i64 %indvars.iv, i64 265 store i32 %add31, ptr %arrayidx34, align 866 %add38 = add nsw i32 %add24, %add3167 %12 = or disjoint i64 %0, 368 %arrayidx41 = getelementptr inbounds i32, ptr %diff, i64 %1269 %13 = load i32, ptr %arrayidx41, align 470 %14 = or disjoint i64 %0, 771 %arrayidx44 = getelementptr inbounds i32, ptr %diff, i64 %1472 %15 = load i32, ptr %arrayidx44, align 473 %add45 = add nsw i32 %15, %1374 %arrayidx48 = getelementptr inbounds [8 x [8 x i32]], ptr %m2, i64 0, i64 %indvars.iv, i64 375 store i32 %add45, ptr %arrayidx48, align 476 %add52 = add nsw i32 %add38, %add4577 %indvars.iv.next = add nuw nsw i64 %indvars.iv, 178 %exitcond = icmp eq i64 %indvars.iv.next, 879 br i1 %exitcond, label %for.end, label %for.body80 81for.end: ; preds = %for.body82 call void @ff(ptr %m2) #183 ret i32 %add5284}85 86declare void @ff(ptr) #287 88 89