164 lines · plain
1; NOTE: Assertions have been autogenerated by utils/update_test_checks.py2; RUN: opt -passes=loop-unroll -mtriple=thumbv7a-unknown-linux-gnueabihf -S %s | FileCheck %s3 4; Check we unroll even with optsize, if the result is smaller, either because5; we have single iteration loops or bodies with constant folding opportunities6; after fully unrolling.7 8; TODO: Looks like we should enable some unrolling for M-class, even when9; optimising for size.10 11declare i32 @get()12 13define void @fully_unrolled_single_iteration(ptr %src) #0 {14; CHECK-LABEL: @fully_unrolled_single_iteration(15; CHECK-NEXT: entry:16; CHECK-NEXT: [[ARR:%.*]] = alloca [4 x i32], align 417; CHECK-NEXT: br label [[FOR_BODY:%.*]]18; CHECK: for.body:19; CHECK-NEXT: [[V:%.*]] = load i32, ptr [[SRC:%.*]]20; CHECK-NEXT: store i32 [[V]], ptr [[ARR]], align 421; CHECK-NEXT: call void @use(ptr nonnull [[ARR]])22; CHECK-NEXT: ret void23;24entry:25 %arr = alloca [4 x i32], align 426 br label %for.body27 28for.body: ; preds = %for.body, %entry29 %indvars.iv = phi i64 [ %indvars.iv.next, %for.body ], [ 0, %entry ]30 %src.idx = getelementptr inbounds i32, ptr %src, i64 %indvars.iv31 %v = load i32, ptr %src.idx32 %arrayidx = getelementptr inbounds [4 x i32], ptr %arr, i64 0, i64 %indvars.iv33 store i32 %v, ptr %arrayidx, align 434 %indvars.iv.next = add nuw nsw i64 %indvars.iv, 135 %exitcond = icmp eq i64 %indvars.iv.next, 136 br i1 %exitcond, label %for.cond.cleanup, label %for.body37 38for.cond.cleanup: ; preds = %for.cond39 call void @use(ptr nonnull %arr) #440 ret void41}42 43 44define void @fully_unrolled_smaller() #0 {45; CHECK-LABEL: @fully_unrolled_smaller(46; CHECK-NEXT: entry:47; CHECK-NEXT: [[ARR:%.*]] = alloca [4 x i32], align 448; CHECK-NEXT: br label [[FOR_BODY:%.*]]49; CHECK: for.body:50; CHECK-NEXT: store i32 16, ptr [[ARR]], align 451; CHECK-NEXT: [[ARRAYIDX_1:%.*]] = getelementptr inbounds [4 x i32], ptr [[ARR]], i64 0, i64 152; CHECK-NEXT: store i32 4104, ptr [[ARRAYIDX_1]], align 453; CHECK-NEXT: [[ARRAYIDX_2:%.*]] = getelementptr inbounds [4 x i32], ptr [[ARR]], i64 0, i64 254; CHECK-NEXT: store i32 1048592, ptr [[ARRAYIDX_2]], align 455; CHECK-NEXT: [[ARRAYIDX_3:%.*]] = getelementptr inbounds [4 x i32], ptr [[ARR]], i64 0, i64 356; CHECK-NEXT: store i32 268435480, ptr [[ARRAYIDX_3]], align 457; CHECK-NEXT: call void @use(ptr nonnull [[ARR]])58; CHECK-NEXT: ret void59;60entry:61 %arr = alloca [4 x i32], align 462 br label %for.body63 64for.body: ; preds = %for.body, %entry65 %indvars.iv = phi i64 [ %indvars.iv.next, %for.body ], [ 0, %entry ]66 %indvars.iv.tr = trunc i64 %indvars.iv to i3267 %shl.0 = shl i32 %indvars.iv.tr, 368 %shl.1 = shl i32 16, %shl.069 %or = or i32 %shl.1, %shl.070 %arrayidx = getelementptr inbounds [4 x i32], ptr %arr, i64 0, i64 %indvars.iv71 store i32 %or, ptr %arrayidx, align 472 %indvars.iv.next = add nuw nsw i64 %indvars.iv, 173 %exitcond = icmp eq i64 %indvars.iv, 374 br i1 %exitcond, label %for.cond.cleanup, label %for.body75 76for.cond.cleanup: ; preds = %for.cond77 call void @use(ptr nonnull %arr) #478 ret void79}80 81define void @fully_unrolled_smaller_Oz() #1 {82; CHECK-LABEL: @fully_unrolled_smaller_Oz(83; CHECK-NEXT: entry:84; CHECK-NEXT: [[ARR:%.*]] = alloca [4 x i32], align 485; CHECK-NEXT: br label [[FOR_BODY:%.*]]86; CHECK: for.body:87; CHECK-NEXT: store i32 16, ptr [[ARR]], align 488; CHECK-NEXT: [[ARRAYIDX_1:%.*]] = getelementptr inbounds [4 x i32], ptr [[ARR]], i64 0, i64 189; CHECK-NEXT: store i32 4104, ptr [[ARRAYIDX_1]], align 490; CHECK-NEXT: [[ARRAYIDX_2:%.*]] = getelementptr inbounds [4 x i32], ptr [[ARR]], i64 0, i64 291; CHECK-NEXT: store i32 1048592, ptr [[ARRAYIDX_2]], align 492; CHECK-NEXT: [[ARRAYIDX_3:%.*]] = getelementptr inbounds [4 x i32], ptr [[ARR]], i64 0, i64 393; CHECK-NEXT: store i32 268435480, ptr [[ARRAYIDX_3]], align 494; CHECK-NEXT: call void @use(ptr nonnull [[ARR]])95; CHECK-NEXT: ret void96;97entry:98 %arr = alloca [4 x i32], align 499 br label %for.body100 101for.body: ; preds = %for.body, %entry102 %indvars.iv = phi i64 [ %indvars.iv.next, %for.body ], [ 0, %entry ]103 %indvars.iv.tr = trunc i64 %indvars.iv to i32104 %shl.0 = shl i32 %indvars.iv.tr, 3105 %shl.1 = shl i32 16, %shl.0106 %or = or i32 %shl.1, %shl.0107 %arrayidx = getelementptr inbounds [4 x i32], ptr %arr, i64 0, i64 %indvars.iv108 store i32 %or, ptr %arrayidx, align 4109 %indvars.iv.next = add nuw nsw i64 %indvars.iv, 1110 %exitcond = icmp eq i64 %indvars.iv, 3111 br i1 %exitcond, label %for.cond.cleanup, label %for.body112 113for.cond.cleanup: ; preds = %for.cond114 call void @use(ptr nonnull %arr) #4115 ret void116}117 118 119define void @fully_unrolled_bigger() #0 {120; CHECK-LABEL: @fully_unrolled_bigger(121; CHECK-NEXT: entry:122; CHECK-NEXT: [[ARR:%.*]] = alloca [4 x i32], align 4123; CHECK-NEXT: br label [[FOR_BODY:%.*]]124; CHECK: for.body:125; CHECK-NEXT: [[INDVARS_IV:%.*]] = phi i64 [ [[INDVARS_IV_NEXT:%.*]], [[FOR_BODY]] ], [ 0, [[ENTRY:%.*]] ]126; CHECK-NEXT: [[INDVARS_IV_TR:%.*]] = trunc i64 [[INDVARS_IV]] to i32127; CHECK-NEXT: [[SHL_0:%.*]] = shl i32 [[INDVARS_IV_TR]], 3128; CHECK-NEXT: [[SHL_1:%.*]] = shl i32 16, [[SHL_0]]129; CHECK-NEXT: [[OR:%.*]] = or i32 [[SHL_1]], [[SHL_0]]130; CHECK-NEXT: [[ARRAYIDX:%.*]] = getelementptr inbounds [4 x i32], ptr [[ARR]], i64 0, i64 [[INDVARS_IV]]131; CHECK-NEXT: store i32 [[OR]], ptr [[ARRAYIDX]], align 4132; CHECK-NEXT: [[INDVARS_IV_NEXT]] = add nuw nsw i64 [[INDVARS_IV]], 1133; CHECK-NEXT: [[EXITCOND:%.*]] = icmp eq i64 [[INDVARS_IV]], 7134; CHECK-NEXT: br i1 [[EXITCOND]], label [[FOR_COND_CLEANUP:%.*]], label [[FOR_BODY]]135; CHECK: for.cond.cleanup:136; CHECK-NEXT: call void @use(ptr nonnull [[ARR]])137; CHECK-NEXT: ret void138;139entry:140 %arr = alloca [4 x i32], align 4141 br label %for.body142 143for.body: ; preds = %for.body, %entry144 %indvars.iv = phi i64 [ %indvars.iv.next, %for.body ], [ 0, %entry ]145 %indvars.iv.tr = trunc i64 %indvars.iv to i32146 %shl.0 = shl i32 %indvars.iv.tr, 3147 %shl.1 = shl i32 16, %shl.0148 %or = or i32 %shl.1, %shl.0149 %arrayidx = getelementptr inbounds [4 x i32], ptr %arr, i64 0, i64 %indvars.iv150 store i32 %or, ptr %arrayidx, align 4151 %indvars.iv.next = add nuw nsw i64 %indvars.iv, 1152 %exitcond = icmp eq i64 %indvars.iv, 7153 br i1 %exitcond, label %for.cond.cleanup, label %for.body154 155for.cond.cleanup: ; preds = %for.cond156 call void @use(ptr nonnull %arr) #4157 ret void158}159 160declare void @use(ptr)161 162attributes #0 = { optsize }163attributes #1 = { minsize optsize }164