106 lines · plain
1; NOTE: Assertions have been autogenerated by utils/update_llc_test_checks.py2; RUN: llc -mtriple=thumbv8.1m.main-none-none-eabi -mattr=+mve -verify-machineinstrs %s -o - | FileCheck %s3 4; This test has an instruction that gets sunk into the loop, that is a5; active.lane.mask operand. (%exitcount.ptrcnt.to.int = ptrtoint). We6; need to make sure it is loop invariant.7 8define i32 @a(ptr readnone %b, ptr %c) {9; CHECK-LABEL: a:10; CHECK: @ %bb.0: @ %entry11; CHECK-NEXT: .save {r4, lr}12; CHECK-NEXT: push {r4, lr}13; CHECK-NEXT: cmp r0, r114; CHECK-NEXT: it ls15; CHECK-NEXT: popls {r4, pc}16; CHECK-NEXT: .LBB0_1: @ %while.body.preheader17; CHECK-NEXT: subs r4, r0, r118; CHECK-NEXT: movs r2, #019; CHECK-NEXT: mov r3, r120; CHECK-NEXT: dlstp.8 lr, r421; CHECK-NEXT: .LBB0_2: @ %vector.body22; CHECK-NEXT: @ =>This Inner Loop Header: Depth=123; CHECK-NEXT: adds r0, r1, r224; CHECK-NEXT: vidup.u8 q0, r0, #125; CHECK-NEXT: adds r2, #1626; CHECK-NEXT: vstrb.8 q0, [r3], #1627; CHECK-NEXT: letp lr, .LBB0_228; CHECK-NEXT: @ %bb.3: @ %while.end29; CHECK-NEXT: pop {r4, pc}30entry:31 %cmp3 = icmp ugt ptr %b, %c32 br i1 %cmp3, label %while.body.preheader, label %while.end33 34while.body.preheader: ; preds = %entry35 %c5 = ptrtoint ptr %c to i3236 %0 = sub i32 0, %c537 %uglygep = getelementptr i8, ptr %b, i32 %038 %exitcount.ptrcnt.to.int = ptrtoint ptr %uglygep to i3239 %n.rnd.up = add i32 %exitcount.ptrcnt.to.int, 1540 %n.vec = and i32 %n.rnd.up, -1641 br label %vector.body42 43vector.body: ; preds = %vector.body, %while.body.preheader44 %index = phi i32 [ 0, %while.body.preheader ], [ %index.next, %vector.body ]45 %next.gep = getelementptr i8, ptr %c, i32 %index46 %1 = or disjoint i32 %index, 147 %next.gep7 = getelementptr i8, ptr %c, i32 %148 %2 = or disjoint i32 %index, 249 %next.gep8 = getelementptr i8, ptr %c, i32 %250 %3 = or disjoint i32 %index, 351 %next.gep9 = getelementptr i8, ptr %c, i32 %352 %4 = or disjoint i32 %index, 453 %next.gep10 = getelementptr i8, ptr %c, i32 %454 %5 = or disjoint i32 %index, 555 %next.gep11 = getelementptr i8, ptr %c, i32 %556 %6 = or disjoint i32 %index, 657 %next.gep12 = getelementptr i8, ptr %c, i32 %658 %7 = or disjoint i32 %index, 759 %next.gep13 = getelementptr i8, ptr %c, i32 %760 %8 = or disjoint i32 %index, 861 %next.gep14 = getelementptr i8, ptr %c, i32 %862 %9 = or disjoint i32 %index, 963 %next.gep15 = getelementptr i8, ptr %c, i32 %964 %10 = or disjoint i32 %index, 1065 %next.gep16 = getelementptr i8, ptr %c, i32 %1066 %11 = or disjoint i32 %index, 1167 %next.gep17 = getelementptr i8, ptr %c, i32 %1168 %12 = or disjoint i32 %index, 1269 %next.gep18 = getelementptr i8, ptr %c, i32 %1270 %13 = or disjoint i32 %index, 1371 %next.gep19 = getelementptr i8, ptr %c, i32 %1372 %14 = or disjoint i32 %index, 1473 %next.gep20 = getelementptr i8, ptr %c, i32 %1474 %15 = or disjoint i32 %index, 1575 %next.gep21 = getelementptr i8, ptr %c, i32 %1576 %16 = insertelement <16 x ptr> poison, ptr %next.gep, i32 077 %17 = insertelement <16 x ptr> %16, ptr %next.gep7, i32 178 %18 = insertelement <16 x ptr> %17, ptr %next.gep8, i32 279 %19 = insertelement <16 x ptr> %18, ptr %next.gep9, i32 380 %20 = insertelement <16 x ptr> %19, ptr %next.gep10, i32 481 %21 = insertelement <16 x ptr> %20, ptr %next.gep11, i32 582 %22 = insertelement <16 x ptr> %21, ptr %next.gep12, i32 683 %23 = insertelement <16 x ptr> %22, ptr %next.gep13, i32 784 %24 = insertelement <16 x ptr> %23, ptr %next.gep14, i32 885 %25 = insertelement <16 x ptr> %24, ptr %next.gep15, i32 986 %26 = insertelement <16 x ptr> %25, ptr %next.gep16, i32 1087 %27 = insertelement <16 x ptr> %26, ptr %next.gep17, i32 1188 %28 = insertelement <16 x ptr> %27, ptr %next.gep18, i32 1289 %29 = insertelement <16 x ptr> %28, ptr %next.gep19, i32 1390 %30 = insertelement <16 x ptr> %29, ptr %next.gep20, i32 1491 %31 = insertelement <16 x ptr> %30, ptr %next.gep21, i32 1592 %active.lane.mask = call <16 x i1> @llvm.get.active.lane.mask.v16i1.i32(i32 %index, i32 %exitcount.ptrcnt.to.int)93 %32 = ptrtoint <16 x ptr> %31 to <16 x i32>94 %33 = trunc <16 x i32> %32 to <16 x i8>95 call void @llvm.masked.store.v16i8.p0(<16 x i8> %33, ptr %next.gep, i32 1, <16 x i1> %active.lane.mask)96 %index.next = add i32 %index, 1697 %34 = icmp eq i32 %index.next, %n.vec98 br i1 %34, label %while.end, label %vector.body99 100while.end: ; preds = %vector.body, %entry101 ret i32 undef102}103 104declare <16 x i1> @llvm.get.active.lane.mask.v16i1.i32(i32, i32)105declare void @llvm.masked.store.v16i8.p0(<16 x i8>, ptr, i32 immarg, <16 x i1>)106