531 lines · plain
1; NOTE: Assertions have been autogenerated by utils/update_llc_test_checks.py UTC_ARGS: --version 32; RUN: llc -mtriple=aarch64-linux-gnu < %s | FileCheck %s3 4; One dimensional loop with load that can be hoisted outside of loop5; for (int i = 0; i < N; ++i)6; if (!memcmp(a[i], b, 4))7; sum += 1;8;9define i64 @one_dimensional(ptr %a, ptr %b, i64 %N) {10; CHECK-LABEL: one_dimensional:11; CHECK: // %bb.0: // %entry12; CHECK-NEXT: ldr w9, [x1]13; CHECK-NEXT: mov x8, xzr14; CHECK-NEXT: .LBB0_1: // %for.body15; CHECK-NEXT: // =>This Inner Loop Header: Depth=116; CHECK-NEXT: ldr x10, [x0], #817; CHECK-NEXT: ldr w10, [x10]18; CHECK-NEXT: cmp w10, w919; CHECK-NEXT: cinc x8, x8, eq20; CHECK-NEXT: subs x2, x2, #121; CHECK-NEXT: b.ne .LBB0_122; CHECK-NEXT: // %bb.2: // %for.exit23; CHECK-NEXT: mov x0, x824; CHECK-NEXT: ret25entry:26 br label %for.body27 28for.body: ; preds = %entry, %for.body29 %i.06 = phi i64 [ %inc, %for.body ], [ 0, %entry ]30 %sum.05 = phi i64 [ %spec.select, %for.body ], [ 0, %entry ]31 %arrayidx = getelementptr inbounds ptr, ptr %a, i64 %i.0632 %0 = load ptr, ptr %arrayidx, align 833 %bcmp = tail call i32 @bcmp(ptr %0, ptr %b, i64 4)34 %tobool = icmp eq i32 %bcmp, 035 %add = zext i1 %tobool to i6436 %spec.select = add i64 %sum.05, %add37 %inc = add nuw i64 %i.06, 138 %exitcond = icmp eq i64 %inc, %N39 br i1 %exitcond, label %for.exit, label %for.body40 41for.exit: ; preds = %for.body42 ret i64 %spec.select43}44 45; Same but loop is two dimensional. Load is hosted outside of both loops46; for (int i = 0; i < N; ++i)47; for (int j = 0; j < M; ++j)48; if (!memcmp(a[i][j], b, 4))49; sum += 1;50;51define i64 @two_dimensional(ptr %a, ptr %b, i64 %N, i64 %M) {52; CHECK-LABEL: two_dimensional:53; CHECK: // %bb.0: // %entry54; CHECK-NEXT: ldr w10, [x1]55; CHECK-NEXT: mov x9, xzr56; CHECK-NEXT: mov x8, xzr57; CHECK-NEXT: .LBB1_1: // %for.cond1.preheader58; CHECK-NEXT: // =>This Loop Header: Depth=159; CHECK-NEXT: // Child Loop BB1_2 Depth 260; CHECK-NEXT: ldr x11, [x0, x9, lsl #3]61; CHECK-NEXT: mov x12, x362; CHECK-NEXT: .LBB1_2: // %for.body463; CHECK-NEXT: // Parent Loop BB1_1 Depth=164; CHECK-NEXT: // => This Inner Loop Header: Depth=265; CHECK-NEXT: ldr x13, [x11], #866; CHECK-NEXT: ldr w13, [x13]67; CHECK-NEXT: cmp w13, w1068; CHECK-NEXT: cinc x8, x8, eq69; CHECK-NEXT: subs x12, x12, #170; CHECK-NEXT: b.ne .LBB1_271; CHECK-NEXT: // %bb.3: // %for.cond1.for.exit3_crit_edge72; CHECK-NEXT: // in Loop: Header=BB1_1 Depth=173; CHECK-NEXT: add x9, x9, #174; CHECK-NEXT: cmp x9, x275; CHECK-NEXT: b.ne .LBB1_176; CHECK-NEXT: // %bb.4: // %for.exit77; CHECK-NEXT: mov x0, x878; CHECK-NEXT: ret79entry:80 br label %for.cond1.preheader81 82for.cond1.preheader: ; preds = %entry, %for.cond1.for.exit3_crit_edge83 %i.019 = phi i64 [ %inc7, %for.cond1.for.exit3_crit_edge ], [ 0, %entry ]84 %sum.018 = phi i64 [ %spec.select, %for.cond1.for.exit3_crit_edge ], [ 0, %entry ]85 %arrayidx = getelementptr inbounds ptr, ptr %a, i64 %i.01986 %0 = load ptr, ptr %arrayidx, align 887 br label %for.body488 89for.body4: ; preds = %for.cond1.preheader, %for.body490 %j.016 = phi i64 [ 0, %for.cond1.preheader ], [ %inc, %for.body4 ]91 %sum.115 = phi i64 [ %sum.018, %for.cond1.preheader ], [ %spec.select, %for.body4 ]92 %arrayidx5 = getelementptr inbounds ptr, ptr %0, i64 %j.01693 %1 = load ptr, ptr %arrayidx5, align 894 %bcmp = tail call i32 @bcmp(ptr %1, ptr %b, i64 4)95 %tobool = icmp eq i32 %bcmp, 096 %add = zext i1 %tobool to i6497 %spec.select = add i64 %sum.115, %add98 %inc = add nuw i64 %j.016, 199 %exitcond = icmp eq i64 %inc, %M100 br i1 %exitcond, label %for.cond1.for.exit3_crit_edge, label %for.body4101 102for.cond1.for.exit3_crit_edge: ; preds = %for.body4103 %inc7 = add nuw i64 %i.019, 1104 %exitcond22 = icmp eq i64 %inc7, %N105 br i1 %exitcond22, label %for.exit, label %for.cond1.preheader106 107for.exit: ; preds = %for.cond1.for.exit3_crit_edge108 ret i64 %spec.select109}110 111; Same but loop is three dimensional. Load is hosted outside of all three loops112; for (int i = 0; i < N; ++i)113; for (int j = 0; j < M; ++j)114; for (int k = 0; k < K; ++k)115; if (!memcmp(a[i][j][k], b, 4))116; sum += 1;117;118define i64 @three_dimensional(ptr %a, ptr %b, i64 %N, i64 %M, i64 %K) {119; CHECK-LABEL: three_dimensional:120; CHECK: // %bb.0: // %entry121; CHECK-NEXT: ldr w10, [x1]122; CHECK-NEXT: mov x9, xzr123; CHECK-NEXT: mov x8, xzr124; CHECK-NEXT: .LBB2_1: // %for.cond1.preheader125; CHECK-NEXT: // =>This Loop Header: Depth=1126; CHECK-NEXT: // Child Loop BB2_2 Depth 2127; CHECK-NEXT: // Child Loop BB2_3 Depth 3128; CHECK-NEXT: ldr x11, [x0, x9, lsl #3]129; CHECK-NEXT: mov x12, xzr130; CHECK-NEXT: .LBB2_2: // %for.cond5.preheader131; CHECK-NEXT: // Parent Loop BB2_1 Depth=1132; CHECK-NEXT: // => This Loop Header: Depth=2133; CHECK-NEXT: // Child Loop BB2_3 Depth 3134; CHECK-NEXT: ldr x13, [x11, x12, lsl #3]135; CHECK-NEXT: mov x14, x4136; CHECK-NEXT: .LBB2_3: // %for.body8137; CHECK-NEXT: // Parent Loop BB2_1 Depth=1138; CHECK-NEXT: // Parent Loop BB2_2 Depth=2139; CHECK-NEXT: // => This Inner Loop Header: Depth=3140; CHECK-NEXT: ldr x15, [x13], #8141; CHECK-NEXT: ldr w15, [x15]142; CHECK-NEXT: cmp w15, w10143; CHECK-NEXT: cinc x8, x8, eq144; CHECK-NEXT: subs x14, x14, #1145; CHECK-NEXT: b.ne .LBB2_3146; CHECK-NEXT: // %bb.4: // %for.cond5.for.cond147; CHECK-NEXT: // in Loop: Header=BB2_2 Depth=2148; CHECK-NEXT: add x12, x12, #1149; CHECK-NEXT: cmp x12, x3150; CHECK-NEXT: b.ne .LBB2_2151; CHECK-NEXT: // %bb.5: // %for.cond1.for.cond152; CHECK-NEXT: // in Loop: Header=BB2_1 Depth=1153; CHECK-NEXT: add x9, x9, #1154; CHECK-NEXT: cmp x9, x2155; CHECK-NEXT: b.ne .LBB2_1156; CHECK-NEXT: // %bb.6: // %for.exit157; CHECK-NEXT: mov x0, x8158; CHECK-NEXT: ret159entry:160 br label %for.cond1.preheader161 162for.cond1.preheader: ; preds = %entry, %for.cond1.for.cond163 %i.033 = phi i64 [ %inc15, %for.cond1.for.cond ], [ 0, %entry ]164 %sum.032 = phi i64 [ %spec.select, %for.cond1.for.cond ], [ 0, %entry ]165 %arrayidx = getelementptr inbounds ptr, ptr %a, i64 %i.033166 %0 = load ptr, ptr %arrayidx, align 8167 br label %for.cond5.preheader168 169for.cond5.preheader: ; preds = %for.cond5.for.cond, %for.cond1.preheader170 %j.029 = phi i64 [ 0, %for.cond1.preheader ], [ %inc12, %for.cond5.for.cond ]171 %sum.128 = phi i64 [ %sum.032, %for.cond1.preheader ], [ %spec.select, %for.cond5.for.cond ]172 %arrayidx9 = getelementptr inbounds ptr, ptr %0, i64 %j.029173 %1 = load ptr, ptr %arrayidx9, align 8174 br label %for.body8175 176for.body8: ; preds = %for.body8, %for.cond5.preheader177 %k.026 = phi i64 [ 0, %for.cond5.preheader ], [ %inc, %for.body8 ]178 %sum.225 = phi i64 [ %sum.128, %for.cond5.preheader ], [ %spec.select, %for.body8 ]179 %arrayidx10 = getelementptr inbounds ptr, ptr %1, i64 %k.026180 %2 = load ptr, ptr %arrayidx10, align 8181 %bcmp = tail call i32 @bcmp(ptr %2, ptr %b, i64 4)182 %tobool = icmp eq i32 %bcmp, 0183 %add = zext i1 %tobool to i64184 %spec.select = add i64 %sum.225, %add185 %inc = add nuw i64 %k.026, 1186 %exitcond = icmp eq i64 %inc, %K187 br i1 %exitcond, label %for.cond5.for.cond, label %for.body8188 189for.cond5.for.cond: ; preds = %for.body8190 %inc12 = add nuw i64 %j.029, 1191 %exitcond44 = icmp eq i64 %inc12, %M192 br i1 %exitcond44, label %for.cond1.for.cond, label %for.cond5.preheader193 194for.cond1.for.cond: ; preds = %for.cond5.for.cond195 %inc15 = add nuw i64 %i.033, 1196 %exitcond45 = icmp eq i64 %inc15, %N197 br i1 %exitcond45, label %for.exit, label %for.cond1.preheader198 199for.exit: ; preds = %for.cond1.for.cond200 ret i64 %spec.select201}202 203; Three dimensional loop but `b` is invariant only relatively to the inner loop.204; Make sure that load is hoisted only outside of first loop205; for (int i = 0; i < N; ++i)206; for (int j = 0; j < M; ++j)207; for (int k = 0; k < K; ++k)208; if (!memcmp(a[i][j][k], b[j], 4))209; sum += 1;210;211define i64 @three_dimensional_middle(ptr %a, ptr %b, i64 %N, i64 %M, i64 %K) {212; CHECK-LABEL: three_dimensional_middle:213; CHECK: // %bb.0: // %entry214; CHECK-NEXT: mov x9, xzr215; CHECK-NEXT: mov x8, xzr216; CHECK-NEXT: .LBB3_1: // %for.cond1.preheader217; CHECK-NEXT: // =>This Loop Header: Depth=1218; CHECK-NEXT: // Child Loop BB3_2 Depth 2219; CHECK-NEXT: // Child Loop BB3_3 Depth 3220; CHECK-NEXT: ldr x10, [x0, x9, lsl #3]221; CHECK-NEXT: mov x11, xzr222; CHECK-NEXT: .LBB3_2: // %for.cond5.preheader223; CHECK-NEXT: // Parent Loop BB3_1 Depth=1224; CHECK-NEXT: // => This Loop Header: Depth=2225; CHECK-NEXT: // Child Loop BB3_3 Depth 3226; CHECK-NEXT: ldr x13, [x1, x11, lsl #3]227; CHECK-NEXT: ldr x12, [x10, x11, lsl #3]228; CHECK-NEXT: mov x14, x4229; CHECK-NEXT: ldr w13, [x13]230; CHECK-NEXT: .LBB3_3: // %for.body8231; CHECK-NEXT: // Parent Loop BB3_1 Depth=1232; CHECK-NEXT: // Parent Loop BB3_2 Depth=2233; CHECK-NEXT: // => This Inner Loop Header: Depth=3234; CHECK-NEXT: ldr x15, [x12], #8235; CHECK-NEXT: ldr w15, [x15]236; CHECK-NEXT: cmp w15, w13237; CHECK-NEXT: cinc x8, x8, eq238; CHECK-NEXT: subs x14, x14, #1239; CHECK-NEXT: b.ne .LBB3_3240; CHECK-NEXT: // %bb.4: // %for.cond5.for.cond241; CHECK-NEXT: // in Loop: Header=BB3_2 Depth=2242; CHECK-NEXT: add x11, x11, #1243; CHECK-NEXT: cmp x11, x3244; CHECK-NEXT: b.ne .LBB3_2245; CHECK-NEXT: // %bb.5: // %for.cond1.for.cond246; CHECK-NEXT: // in Loop: Header=BB3_1 Depth=1247; CHECK-NEXT: add x9, x9, #1248; CHECK-NEXT: cmp x9, x2249; CHECK-NEXT: b.ne .LBB3_1250; CHECK-NEXT: // %bb.6: // %for.exit251; CHECK-NEXT: mov x0, x8252; CHECK-NEXT: ret253entry:254 br label %for.cond1.preheader255 256for.cond1.preheader: ; preds = %entry, %for.cond1.for.cond257 %i.035 = phi i64 [ %inc16, %for.cond1.for.cond ], [ 0, %entry ]258 %sum.034 = phi i64 [ %spec.select, %for.cond1.for.cond ], [ 0, %entry ]259 %arrayidx = getelementptr inbounds ptr, ptr %a, i64 %i.035260 %0 = load ptr, ptr %arrayidx, align 8261 br label %for.cond5.preheader262 263for.cond5.preheader: ; preds = %for.cond5.for.cond, %for.cond1.preheader264 %j.031 = phi i64 [ 0, %for.cond1.preheader ], [ %inc13, %for.cond5.for.cond ]265 %sum.130 = phi i64 [ %sum.034, %for.cond1.preheader ], [ %spec.select, %for.cond5.for.cond ]266 %arrayidx9 = getelementptr inbounds ptr, ptr %0, i64 %j.031267 %1 = load ptr, ptr %arrayidx9, align 8268 %arrayidx11 = getelementptr inbounds ptr, ptr %b, i64 %j.031269 %2 = load ptr, ptr %arrayidx11, align 8270 br label %for.body8271 272for.body8: ; preds = %for.body8, %for.cond5.preheader273 %k.028 = phi i64 [ 0, %for.cond5.preheader ], [ %inc, %for.body8 ]274 %sum.227 = phi i64 [ %sum.130, %for.cond5.preheader ], [ %spec.select, %for.body8 ]275 %arrayidx10 = getelementptr inbounds ptr, ptr %1, i64 %k.028276 %3 = load ptr, ptr %arrayidx10, align 8277 %bcmp = tail call i32 @bcmp(ptr %3, ptr %2, i64 4)278 %tobool = icmp eq i32 %bcmp, 0279 %add = zext i1 %tobool to i64280 %spec.select = add i64 %sum.227, %add281 %inc = add nuw i64 %k.028, 1282 %exitcond = icmp eq i64 %inc, %K283 br i1 %exitcond, label %for.cond5.for.cond, label %for.body8284 285for.cond5.for.cond: ; preds = %for.body8286 %inc13 = add nuw i64 %j.031, 1287 %exitcond46 = icmp eq i64 %inc13, %M288 br i1 %exitcond46, label %for.cond1.for.cond, label %for.cond5.preheader289 290for.cond1.for.cond: ; preds = %for.cond5.for.cond291 %inc16 = add nuw i64 %i.035, 1292 %exitcond47 = icmp eq i64 %inc16, %N293 br i1 %exitcond47, label %for.exit, label %for.cond1.preheader294 295for.exit: ; preds = %for.cond1.for.cond296 ret i64 %spec.select297}298 299; Make sure that store inside loop prevents hoisting invariant loads300; for (int i = 0; i < N; ++i)301; c[i] = memcmp(a[i], b, 4);302;303define void @one_dimensional_with_store(ptr %a, ptr %b, ptr %c, i32 %N) {304; CHECK-LABEL: one_dimensional_with_store:305; CHECK: // %bb.0: // %entry306; CHECK-NEXT: mov w8, w3307; CHECK-NEXT: .LBB4_1: // %for.body308; CHECK-NEXT: // =>This Inner Loop Header: Depth=1309; CHECK-NEXT: ldr x9, [x0], #8310; CHECK-NEXT: ldr w10, [x1]311; CHECK-NEXT: ldr w9, [x9]312; CHECK-NEXT: rev w10, w10313; CHECK-NEXT: rev w9, w9314; CHECK-NEXT: cmp w9, w10315; CHECK-NEXT: cset w9, hi316; CHECK-NEXT: csinv w9, w9, wzr, hs317; CHECK-NEXT: subs x8, x8, #1318; CHECK-NEXT: strb w9, [x2], #1319; CHECK-NEXT: b.ne .LBB4_1320; CHECK-NEXT: // %bb.2: // %for.exit321; CHECK-NEXT: ret322entry:323 br label %for.body.preheader324 325for.body.preheader: ; preds = %entry326 %wide.trip.count = zext i32 %N to i64327 br label %for.body328 329for.body: ; preds = %for.body.preheader, %for.body330 %indvars.iv = phi i64 [ 0, %for.body.preheader ], [ %indvars.iv.next, %for.body ]331 %arrayidx = getelementptr inbounds ptr, ptr %a, i64 %indvars.iv332 %0 = load ptr, ptr %arrayidx, align 8333 %call = tail call i32 @memcmp(ptr %0, ptr %b, i64 4)334 %conv = trunc i32 %call to i8335 %arrayidx2 = getelementptr inbounds i8, ptr %c, i64 %indvars.iv336 store i8 %conv, ptr %arrayidx2, align 1337 %indvars.iv.next = add nuw nsw i64 %indvars.iv, 1338 %exitcond.not = icmp eq i64 %indvars.iv.next, %wide.trip.count339 br i1 %exitcond.not, label %for.exit, label %for.body340 341for.exit: ; preds = %for.body342 ret void343}344 345; Make sure that call inside loop prevents hoisting invariant loads346;347define i32 @one_dimensional_with_call(ptr %a, ptr %b, i32 %N) {348; CHECK-LABEL: one_dimensional_with_call:349; CHECK: // %bb.0: // %entry350; CHECK-NEXT: str x30, [sp, #-48]! // 8-byte Folded Spill351; CHECK-NEXT: stp x22, x21, [sp, #16] // 16-byte Folded Spill352; CHECK-NEXT: stp x20, x19, [sp, #32] // 16-byte Folded Spill353; CHECK-NEXT: .cfi_def_cfa_offset 48354; CHECK-NEXT: .cfi_offset w19, -8355; CHECK-NEXT: .cfi_offset w20, -16356; CHECK-NEXT: .cfi_offset w21, -24357; CHECK-NEXT: .cfi_offset w22, -32358; CHECK-NEXT: .cfi_offset w30, -48359; CHECK-NEXT: mov x19, x1360; CHECK-NEXT: mov x21, x0361; CHECK-NEXT: mov w20, wzr362; CHECK-NEXT: mov w22, w2363; CHECK-NEXT: .LBB5_1: // %for.body364; CHECK-NEXT: // =>This Inner Loop Header: Depth=1365; CHECK-NEXT: ldr x8, [x21], #8366; CHECK-NEXT: ldr w9, [x19]367; CHECK-NEXT: ldr w8, [x8]368; CHECK-NEXT: cmp w8, w9369; CHECK-NEXT: cinc w20, w20, eq370; CHECK-NEXT: bl func371; CHECK-NEXT: subs x22, x22, #1372; CHECK-NEXT: b.ne .LBB5_1373; CHECK-NEXT: // %bb.2: // %for.exit374; CHECK-NEXT: mov w0, w20375; CHECK-NEXT: ldp x20, x19, [sp, #32] // 16-byte Folded Reload376; CHECK-NEXT: ldp x22, x21, [sp, #16] // 16-byte Folded Reload377; CHECK-NEXT: ldr x30, [sp], #48 // 8-byte Folded Reload378; CHECK-NEXT: ret379entry:380 br label %for.body.preheader381 382for.body.preheader: ; preds = %entry383 %wide.trip.count = zext i32 %N to i64384 br label %for.body385 386for.body: ; preds = %for.body.preheader, %for.body387 %indvars.iv = phi i64 [ 0, %for.body.preheader ], [ %indvars.iv.next, %for.body ]388 %sum.05 = phi i32 [ 0, %for.body.preheader ], [ %spec.select, %for.body ]389 %arrayidx = getelementptr inbounds ptr, ptr %a, i64 %indvars.iv390 %0 = load ptr, ptr %arrayidx, align 8391 %bcmp = tail call i32 @bcmp(ptr %0, ptr %b, i64 4)392 %tobool.not = icmp eq i32 %bcmp, 0393 %add = zext i1 %tobool.not to i32394 %spec.select = add nuw nsw i32 %sum.05, %add395 tail call void @func()396 %indvars.iv.next = add nuw nsw i64 %indvars.iv, 1397 %exitcond.not = icmp eq i64 %indvars.iv.next, %wide.trip.count398 br i1 %exitcond.not, label %for.exit, label %for.body399 400for.exit: ; preds = %for.body401 ret i32 %spec.select402}403 404; One dimensional loop with memcmp size equal six.405; The test shows that shows that several loads can be hoisted at the same time.406; for (int i = 0; i < N; ++i)407; if (!memcmp(a[i], b, 6))408; sum += 1;409;410define i64 @one_dimensional_two_loads(ptr %a, ptr %b, i64 %N) {411; CHECK-LABEL: one_dimensional_two_loads:412; CHECK: // %bb.0: // %entry413; CHECK-NEXT: ldr w9, [x1]414; CHECK-NEXT: ldrh w10, [x1, #4]415; CHECK-NEXT: mov x8, xzr416; CHECK-NEXT: .LBB6_1: // %for.body417; CHECK-NEXT: // =>This Inner Loop Header: Depth=1418; CHECK-NEXT: ldr x11, [x0], #8419; CHECK-NEXT: ldr w12, [x11]420; CHECK-NEXT: ldrh w11, [x11, #4]421; CHECK-NEXT: cmp w12, w9422; CHECK-NEXT: ccmp w11, w10, #0, eq423; CHECK-NEXT: cinc x8, x8, eq424; CHECK-NEXT: subs x2, x2, #1425; CHECK-NEXT: b.ne .LBB6_1426; CHECK-NEXT: // %bb.2: // %for.exit427; CHECK-NEXT: mov x0, x8428; CHECK-NEXT: ret429entry:430 br label %for.body431 432for.body: ; preds = %entry, %for.body433 %i.06 = phi i64 [ %inc, %for.body ], [ 0, %entry ]434 %sum.05 = phi i64 [ %spec.select, %for.body ], [ 0, %entry ]435 %arrayidx = getelementptr inbounds ptr, ptr %a, i64 %i.06436 %0 = load ptr, ptr %arrayidx, align 8437 %bcmp = tail call i32 @bcmp(ptr %0, ptr %b, i64 6)438 %tobool = icmp eq i32 %bcmp, 0439 %add = zext i1 %tobool to i64440 %spec.select = add i64 %sum.05, %add441 %inc = add nuw i64 %i.06, 1442 %exitcond = icmp eq i64 %inc, %N443 br i1 %exitcond, label %for.exit, label %for.body444 445for.exit: ; preds = %for.body446 ret i64 %spec.select447}448 449; See issue https://github.com/llvm/llvm-project/issues/72855450;451; When hoisting instruction out of the loop, ensure that loads are not common452; subexpressions eliminated. In this example pointer %c may alias pointer %b,453; so when hoisting `%y = load i64, ptr %b` instruction we can't replace it with454; `%b.val = load i64, ptr %b`455;456define i64 @hoisting_no_cse(ptr %a, ptr %b, ptr %c, i64 %N) {457; CHECK-LABEL: hoisting_no_cse:458; CHECK: // %bb.0: // %entry459; CHECK-NEXT: ldr x8, [x1]460; CHECK-NEXT: add x8, x8, #1461; CHECK-NEXT: str x8, [x2]462; CHECK-NEXT: mov x8, xzr463; CHECK-NEXT: ldr x9, [x1]464; CHECK-NEXT: .LBB7_1: // %for.body465; CHECK-NEXT: // =>This Inner Loop Header: Depth=1466; CHECK-NEXT: ldr x10, [x0], #8467; CHECK-NEXT: ldr x10, [x10]468; CHECK-NEXT: cmp x10, x9469; CHECK-NEXT: cinc x8, x8, eq470; CHECK-NEXT: subs x3, x3, #1471; CHECK-NEXT: b.ne .LBB7_1472; CHECK-NEXT: // %bb.2: // %for.exit473; CHECK-NEXT: mov x0, x8474; CHECK-NEXT: ret475entry:476 %b.val = load i64, ptr %b477 %b.val.changed = add i64 %b.val, 1478 store i64 %b.val.changed, ptr %c479 br label %for.body480 481for.body: ; preds = %entry, %for.body482 %idx = phi i64 [ %inc, %for.body ], [ 0, %entry ]483 %sum = phi i64 [ %spec.select, %for.body ], [ 0, %entry ]484 %arrayidx = getelementptr inbounds ptr, ptr %a, i64 %idx485 %0 = load ptr, ptr %arrayidx, align 8486 %x = load i64, ptr %0487 %y = load i64, ptr %b488 %cmp = icmp eq i64 %x, %y489 %add = zext i1 %cmp to i64490 %spec.select = add i64 %sum, %add491 %inc = add nuw i64 %idx, 1492 %exitcond = icmp eq i64 %inc, %N493 br i1 %exitcond, label %for.exit, label %for.body494 495for.exit: ; preds = %for.body496 ret i64 %spec.select497}498 499@a = external local_unnamed_addr global i32, align 4500 501; Make sure the load is not hoisted out of the loop across memory barriers.502define i32 @load_between_memory_barriers() {503; CHECK-LABEL: load_between_memory_barriers:504; CHECK: // %bb.0:505; CHECK-NEXT: adrp x8, :got:a506; CHECK-NEXT: ldr x8, [x8, :got_lo12:a]507; CHECK-NEXT: .LBB8_1: // %loop508; CHECK-NEXT: // =>This Inner Loop Header: Depth=1509; CHECK-NEXT: //MEMBARRIER510; CHECK-NEXT: ldr w0, [x8]511; CHECK-NEXT: //MEMBARRIER512; CHECK-NEXT: cbz w0, .LBB8_1513; CHECK-NEXT: // %bb.2: // %exit514; CHECK-NEXT: ret515 br label %loop516 517loop:518 fence syncscope("singlethread") acq_rel519 %l = load i32, ptr @a, align 4520 fence syncscope("singlethread") acq_rel521 %c = icmp eq i32 %l, 0522 br i1 %c, label %loop, label %exit523 524exit:525 ret i32 %l526}527 528declare i32 @bcmp(ptr, ptr, i64)529declare i32 @memcmp(ptr, ptr, i64)530declare void @func()531