197 lines · plain
1; NOTE: Assertions have been autogenerated by utils/update_test_checks.py2; RUN: opt -S -verify-loop-info -passes=irce < %s | FileCheck %s3; RUN: opt -S -verify-loop-info -passes='require<branch-prob>,irce' < %s | FileCheck %s4 5define void @f_0(ptr %arr, ptr %a_len_ptr, i32 %n, ptr %cond_buf) {6; CHECK-LABEL: @f_0(7; CHECK-NEXT: entry:8; CHECK-NEXT: [[LEN:%.*]] = load i32, ptr [[A_LEN_PTR:%.*]], align 4, !range [[RNG0:![0-9]+]]9; CHECK-NEXT: [[FIRST_ITR_CHECK:%.*]] = icmp sgt i32 [[N:%.*]], 010; CHECK-NEXT: br i1 [[FIRST_ITR_CHECK]], label [[LOOP_PREHEADER:%.*]], label [[EXIT:%.*]]11; CHECK: loop.preheader:12; CHECK-NEXT: [[TMP0:%.*]] = add nsw i32 [[LEN]], -413; CHECK-NEXT: [[SMIN:%.*]] = call i32 @llvm.smin.i32(i32 [[N]], i32 [[TMP0]])14; CHECK-NEXT: [[EXIT_MAINLOOP_AT:%.*]] = call i32 @llvm.smax.i32(i32 [[SMIN]], i32 0)15; CHECK-NEXT: [[TMP1:%.*]] = icmp slt i32 0, [[EXIT_MAINLOOP_AT]]16; CHECK-NEXT: br i1 [[TMP1]], label [[LOOP_PREHEADER1:%.*]], label [[MAIN_PSEUDO_EXIT:%.*]]17; CHECK: loop.preheader1:18; CHECK-NEXT: br label [[LOOP:%.*]]19; CHECK: loop:20; CHECK-NEXT: [[IDX:%.*]] = phi i32 [ [[IDX_NEXT:%.*]], [[IN_BOUNDS:%.*]] ], [ 0, [[LOOP_PREHEADER1]] ]21; CHECK-NEXT: [[IDX_NEXT]] = add nsw i32 [[IDX]], 122; CHECK-NEXT: [[IDX_FOR_ABC:%.*]] = add i32 [[IDX]], 423; CHECK-NEXT: [[ABC_ACTUAL:%.*]] = icmp slt i32 [[IDX_FOR_ABC]], [[LEN]]24; CHECK-NEXT: [[COND:%.*]] = load volatile i1, ptr [[COND_BUF:%.*]], align 125; CHECK-NEXT: [[ABC:%.*]] = and i1 [[COND]], true26; CHECK-NEXT: br i1 [[ABC]], label [[IN_BOUNDS]], label [[OUT_OF_BOUNDS_LOOPEXIT2:%.*]], !prof [[PROF1:![0-9]+]]27; CHECK: in.bounds:28; CHECK-NEXT: [[ADDR:%.*]] = getelementptr i32, ptr [[ARR:%.*]], i32 [[IDX_FOR_ABC]]29; CHECK-NEXT: store i32 0, ptr [[ADDR]], align 430; CHECK-NEXT: [[NEXT:%.*]] = icmp slt i32 [[IDX_NEXT]], [[N]]31; CHECK-NEXT: [[TMP2:%.*]] = icmp slt i32 [[IDX_NEXT]], [[EXIT_MAINLOOP_AT]]32; CHECK-NEXT: br i1 [[TMP2]], label [[LOOP]], label [[MAIN_EXIT_SELECTOR:%.*]]33; CHECK: main.exit.selector:34; CHECK-NEXT: [[IDX_NEXT_LCSSA:%.*]] = phi i32 [ [[IDX_NEXT]], [[IN_BOUNDS]] ]35; CHECK-NEXT: [[TMP3:%.*]] = icmp slt i32 [[IDX_NEXT_LCSSA]], [[N]]36; CHECK-NEXT: br i1 [[TMP3]], label [[MAIN_PSEUDO_EXIT]], label [[EXIT_LOOPEXIT:%.*]]37; CHECK: main.pseudo.exit:38; CHECK-NEXT: [[IDX_COPY:%.*]] = phi i32 [ 0, [[LOOP_PREHEADER]] ], [ [[IDX_NEXT_LCSSA]], [[MAIN_EXIT_SELECTOR]] ]39; CHECK-NEXT: [[INDVAR_END:%.*]] = phi i32 [ 0, [[LOOP_PREHEADER]] ], [ [[IDX_NEXT_LCSSA]], [[MAIN_EXIT_SELECTOR]] ]40; CHECK-NEXT: br label [[POSTLOOP:%.*]]41; CHECK: out.of.bounds.loopexit:42; CHECK-NEXT: br label [[OUT_OF_BOUNDS:%.*]]43; CHECK: out.of.bounds.loopexit2:44; CHECK-NEXT: br label [[OUT_OF_BOUNDS]]45; CHECK: out.of.bounds:46; CHECK-NEXT: ret void47; CHECK: exit.loopexit.loopexit:48; CHECK-NEXT: br label [[EXIT_LOOPEXIT]]49; CHECK: exit.loopexit:50; CHECK-NEXT: br label [[EXIT]]51; CHECK: exit:52; CHECK-NEXT: ret void53; CHECK: postloop:54; CHECK-NEXT: br label [[LOOP_POSTLOOP:%.*]]55; CHECK: loop.postloop:56; CHECK-NEXT: [[IDX_POSTLOOP:%.*]] = phi i32 [ [[IDX_NEXT_POSTLOOP:%.*]], [[IN_BOUNDS_POSTLOOP:%.*]] ], [ [[IDX_COPY]], [[POSTLOOP]] ]57; CHECK-NEXT: [[IDX_NEXT_POSTLOOP]] = add i32 [[IDX_POSTLOOP]], 158; CHECK-NEXT: [[IDX_FOR_ABC_POSTLOOP:%.*]] = add i32 [[IDX_POSTLOOP]], 459; CHECK-NEXT: [[ABC_ACTUAL_POSTLOOP:%.*]] = icmp slt i32 [[IDX_FOR_ABC_POSTLOOP]], [[LEN]]60; CHECK-NEXT: [[COND_POSTLOOP:%.*]] = load volatile i1, ptr [[COND_BUF]], align 161; CHECK-NEXT: [[ABC_POSTLOOP:%.*]] = and i1 [[COND_POSTLOOP]], [[ABC_ACTUAL_POSTLOOP]]62; CHECK-NEXT: br i1 [[ABC_POSTLOOP]], label [[IN_BOUNDS_POSTLOOP]], label [[OUT_OF_BOUNDS_LOOPEXIT:%.*]], !prof [[PROF1]]63; CHECK: in.bounds.postloop:64; CHECK-NEXT: [[ADDR_POSTLOOP:%.*]] = getelementptr i32, ptr [[ARR]], i32 [[IDX_FOR_ABC_POSTLOOP]]65; CHECK-NEXT: store i32 0, ptr [[ADDR_POSTLOOP]], align 466; CHECK-NEXT: [[NEXT_POSTLOOP:%.*]] = icmp slt i32 [[IDX_NEXT_POSTLOOP]], [[N]]67; CHECK-NEXT: br i1 [[NEXT_POSTLOOP]], label [[LOOP_POSTLOOP]], label [[EXIT_LOOPEXIT_LOOPEXIT:%.*]], !llvm.loop [[LOOP2:![0-9]+]], !loop_constrainer.loop.clone [[META7:![0-9]+]]68;69entry:70 %len = load i32, ptr %a_len_ptr, !range !071 %first.itr.check = icmp sgt i32 %n, 072 br i1 %first.itr.check, label %loop, label %exit73 74loop:75 %idx = phi i32 [ 0, %entry ] , [ %idx.next, %in.bounds ]76 %idx.next = add i32 %idx, 177 %idx.for.abc = add i32 %idx, 478 %abc.actual = icmp slt i32 %idx.for.abc, %len79 %cond = load volatile i1, ptr %cond_buf80 %abc = and i1 %cond, %abc.actual81 br i1 %abc, label %in.bounds, label %out.of.bounds, !prof !182 83in.bounds:84 %addr = getelementptr i32, ptr %arr, i32 %idx.for.abc85 store i32 0, ptr %addr86 %next = icmp slt i32 %idx.next, %n87 br i1 %next, label %loop, label %exit88 89out.of.bounds:90 ret void91 92exit:93 ret void94}95 96define void @f_1(97; CHECK-LABEL: @f_1(98; CHECK-NEXT: entry:99; CHECK-NEXT: [[LEN_A:%.*]] = load i32, ptr [[A_LEN_PTR:%.*]], align 4, !range [[RNG0]]100; CHECK-NEXT: [[LEN_B:%.*]] = load i32, ptr [[B_LEN_PTR:%.*]], align 4, !range [[RNG0]]101; CHECK-NEXT: [[FIRST_ITR_CHECK:%.*]] = icmp sgt i32 [[N:%.*]], 0102; CHECK-NEXT: br i1 [[FIRST_ITR_CHECK]], label [[LOOP_PREHEADER:%.*]], label [[EXIT:%.*]]103; CHECK: loop.preheader:104; CHECK-NEXT: [[SMIN:%.*]] = call i32 @llvm.smin.i32(i32 [[LEN_B]], i32 [[LEN_A]])105; CHECK-NEXT: [[SMIN1:%.*]] = call i32 @llvm.smin.i32(i32 [[SMIN]], i32 [[N]])106; CHECK-NEXT: [[EXIT_MAINLOOP_AT:%.*]] = call i32 @llvm.smax.i32(i32 [[SMIN1]], i32 0)107; CHECK-NEXT: [[TMP0:%.*]] = icmp slt i32 0, [[EXIT_MAINLOOP_AT]]108; CHECK-NEXT: br i1 [[TMP0]], label [[LOOP_PREHEADER2:%.*]], label [[MAIN_PSEUDO_EXIT:%.*]]109; CHECK: loop.preheader2:110; CHECK-NEXT: br label [[LOOP:%.*]]111; CHECK: loop:112; CHECK-NEXT: [[IDX:%.*]] = phi i32 [ [[IDX_NEXT:%.*]], [[IN_BOUNDS:%.*]] ], [ 0, [[LOOP_PREHEADER2]] ]113; CHECK-NEXT: [[IDX_NEXT]] = add nsw i32 [[IDX]], 1114; CHECK-NEXT: [[ABC_A:%.*]] = icmp slt i32 [[IDX]], [[LEN_A]]115; CHECK-NEXT: [[ABC_B:%.*]] = icmp slt i32 [[IDX]], [[LEN_B]]116; CHECK-NEXT: [[ABC:%.*]] = and i1 true, true117; CHECK-NEXT: br i1 [[ABC]], label [[IN_BOUNDS]], label [[OUT_OF_BOUNDS_LOOPEXIT3:%.*]], !prof [[PROF1]]118; CHECK: in.bounds:119; CHECK-NEXT: [[ADDR_A:%.*]] = getelementptr i32, ptr [[ARR_A:%.*]], i32 [[IDX]]120; CHECK-NEXT: store i32 0, ptr [[ADDR_A]], align 4121; CHECK-NEXT: [[ADDR_B:%.*]] = getelementptr i32, ptr [[ARR_B:%.*]], i32 [[IDX]]122; CHECK-NEXT: store i32 -1, ptr [[ADDR_B]], align 4123; CHECK-NEXT: [[NEXT:%.*]] = icmp slt i32 [[IDX_NEXT]], [[N]]124; CHECK-NEXT: [[TMP1:%.*]] = icmp slt i32 [[IDX_NEXT]], [[EXIT_MAINLOOP_AT]]125; CHECK-NEXT: br i1 [[TMP1]], label [[LOOP]], label [[MAIN_EXIT_SELECTOR:%.*]]126; CHECK: main.exit.selector:127; CHECK-NEXT: [[IDX_NEXT_LCSSA:%.*]] = phi i32 [ [[IDX_NEXT]], [[IN_BOUNDS]] ]128; CHECK-NEXT: [[TMP2:%.*]] = icmp slt i32 [[IDX_NEXT_LCSSA]], [[N]]129; CHECK-NEXT: br i1 [[TMP2]], label [[MAIN_PSEUDO_EXIT]], label [[EXIT_LOOPEXIT:%.*]]130; CHECK: main.pseudo.exit:131; CHECK-NEXT: [[IDX_COPY:%.*]] = phi i32 [ 0, [[LOOP_PREHEADER]] ], [ [[IDX_NEXT_LCSSA]], [[MAIN_EXIT_SELECTOR]] ]132; CHECK-NEXT: [[INDVAR_END:%.*]] = phi i32 [ 0, [[LOOP_PREHEADER]] ], [ [[IDX_NEXT_LCSSA]], [[MAIN_EXIT_SELECTOR]] ]133; CHECK-NEXT: br label [[POSTLOOP:%.*]]134; CHECK: out.of.bounds.loopexit:135; CHECK-NEXT: br label [[OUT_OF_BOUNDS:%.*]]136; CHECK: out.of.bounds.loopexit3:137; CHECK-NEXT: br label [[OUT_OF_BOUNDS]]138; CHECK: out.of.bounds:139; CHECK-NEXT: ret void140; CHECK: exit.loopexit.loopexit:141; CHECK-NEXT: br label [[EXIT_LOOPEXIT]]142; CHECK: exit.loopexit:143; CHECK-NEXT: br label [[EXIT]]144; CHECK: exit:145; CHECK-NEXT: ret void146; CHECK: postloop:147; CHECK-NEXT: br label [[LOOP_POSTLOOP:%.*]]148; CHECK: loop.postloop:149; CHECK-NEXT: [[IDX_POSTLOOP:%.*]] = phi i32 [ [[IDX_NEXT_POSTLOOP:%.*]], [[IN_BOUNDS_POSTLOOP:%.*]] ], [ [[IDX_COPY]], [[POSTLOOP]] ]150; CHECK-NEXT: [[IDX_NEXT_POSTLOOP]] = add i32 [[IDX_POSTLOOP]], 1151; CHECK-NEXT: [[ABC_A_POSTLOOP:%.*]] = icmp slt i32 [[IDX_POSTLOOP]], [[LEN_A]]152; CHECK-NEXT: [[ABC_B_POSTLOOP:%.*]] = icmp slt i32 [[IDX_POSTLOOP]], [[LEN_B]]153; CHECK-NEXT: [[ABC_POSTLOOP:%.*]] = and i1 [[ABC_A_POSTLOOP]], [[ABC_B_POSTLOOP]]154; CHECK-NEXT: br i1 [[ABC_POSTLOOP]], label [[IN_BOUNDS_POSTLOOP]], label [[OUT_OF_BOUNDS_LOOPEXIT:%.*]], !prof [[PROF1]]155; CHECK: in.bounds.postloop:156; CHECK-NEXT: [[ADDR_A_POSTLOOP:%.*]] = getelementptr i32, ptr [[ARR_A]], i32 [[IDX_POSTLOOP]]157; CHECK-NEXT: store i32 0, ptr [[ADDR_A_POSTLOOP]], align 4158; CHECK-NEXT: [[ADDR_B_POSTLOOP:%.*]] = getelementptr i32, ptr [[ARR_B]], i32 [[IDX_POSTLOOP]]159; CHECK-NEXT: store i32 -1, ptr [[ADDR_B_POSTLOOP]], align 4160; CHECK-NEXT: [[NEXT_POSTLOOP:%.*]] = icmp slt i32 [[IDX_NEXT_POSTLOOP]], [[N]]161; CHECK-NEXT: br i1 [[NEXT_POSTLOOP]], label [[LOOP_POSTLOOP]], label [[EXIT_LOOPEXIT_LOOPEXIT:%.*]], !llvm.loop [[LOOP8:![0-9]+]], !loop_constrainer.loop.clone [[META7]]162;163 ptr %arr_a, ptr %a_len_ptr, ptr %arr_b, ptr %b_len_ptr, i32 %n) {164 165 166entry:167 %len.a = load i32, ptr %a_len_ptr, !range !0168 %len.b = load i32, ptr %b_len_ptr, !range !0169 %first.itr.check = icmp sgt i32 %n, 0170 br i1 %first.itr.check, label %loop, label %exit171 172loop:173 %idx = phi i32 [ 0, %entry ] , [ %idx.next, %in.bounds ]174 %idx.next = add i32 %idx, 1175 %abc.a = icmp slt i32 %idx, %len.a176 %abc.b = icmp slt i32 %idx, %len.b177 %abc = and i1 %abc.a, %abc.b178 br i1 %abc, label %in.bounds, label %out.of.bounds, !prof !1179 180in.bounds:181 %addr.a = getelementptr i32, ptr %arr_a, i32 %idx182 store i32 0, ptr %addr.a183 %addr.b = getelementptr i32, ptr %arr_b, i32 %idx184 store i32 -1, ptr %addr.b185 %next = icmp slt i32 %idx.next, %n186 br i1 %next, label %loop, label %exit187 188out.of.bounds:189 ret void190 191exit:192 ret void193}194 195!0 = !{i32 0, i32 2147483647}196!1 = !{!"branch_weights", i32 64, i32 4}197