127 lines · plain
1; Test that loops with sufficient registers do not reload or spill on2; stack. These cases include calls and it is necessary to have the GR128 /3; FP128 registers part of the callee saved registers list in order to avoid4; spilling / reloading.5;6; RUN: llc -switch-peel-threshold=101 < %s -mtriple=s390x-linux-gnu -mcpu=z13 | FileCheck %s7 8%0 = type { ptr, ptr, ptr, i32, ptr, i64, i64, i64, i64, i64, i64, %2, %5, %7 }9%1 = type { i32, i32, ptr, ptr, ptr, ptr, ptr, ptr, ptr }10%2 = type { i64, i64, ptr }11%3 = type { ptr, i64 }12%4 = type { i64, ptr }13%5 = type { i64, i64, ptr }14%6 = type { i64, ptr, i32, i64, ptr }15%7 = type { i64, i64, ptr }16%8 = type { i64, ptr, ptr, ptr, i64, ptr, %5, i32, i64, i64 }17 18declare void @llvm.memcpy.p0.p0.i64(ptr nocapture writeonly, ptr nocapture readonly, i64, i1)19 20define void @fun0(ptr) {21; CHECK-LABEL: .LBB0_422; CHECK: => This Inner Loop Header23; CHECK-NOT: 16-byte Folded Spill24; CHECK-NOT: 16-byte Folded Reload25 26 %2 = load i64, ptr undef, align 827 %3 = udiv i64 128, %228 %4 = mul i64 %3, %229 %5 = load i64, ptr undef, align 830 switch i32 undef, label %36 [31 i32 1, label %632 i32 2, label %733 i32 3, label %834 i32 4, label %935 i32 5, label %1036 i32 6, label %1137 ]38 39; <label>:6: ; preds = %140 br label %1241 42; <label>:7: ; preds = %143 br label %1244 45; <label>:8: ; preds = %146 unreachable47 48; <label>:9: ; preds = %149 unreachable50 51; <label>:10: ; preds = %152 unreachable53 54; <label>:11: ; preds = %155 unreachable56 57; <label>:12: ; preds = %7, %658 %13 = getelementptr inbounds %0, ptr %0, i64 0, i32 559 br label %1460 61; <label>:14: ; preds = %31, %1262 %15 = phi i64 [ undef, %31 ], [ %5, %12 ]63 %16 = phi i64 [ %35, %31 ], [ undef, %12 ]64 %17 = load i64, ptr %13, align 865 %18 = icmp ult i64 %15, %1766 %19 = select i1 %18, i64 %15, i64 %1767 %20 = udiv i64 %19, %468 %21 = icmp ugt i64 %20, 169 %22 = select i1 %21, i64 %20, i64 170 %23 = sub i64 %22, 071 br label %2472 73; <label>:24: ; preds = %24, %1474 %25 = phi i64 [ %23, %14 ], [ %27, %24 ]75 call void @llvm.memcpy.p0.p0.i64(ptr undef, ptr nonnull undef, i64 %4, i1 false)76 %26 = getelementptr inbounds i8, ptr null, i64 %477 store ptr %26, ptr undef, align 878 %27 = add i64 %25, -479 %28 = icmp eq i64 %27, 080 br i1 %28, label %31, label %2481 82; <label>:29: ; preds = %2483 br i1 undef, label %31, label %3084 85; <label>:30: ; preds = %2986 call void @llvm.memcpy.p0.p0.i64(ptr %26, ptr nonnull undef, i64 %4, i1 false)87 br label %3188 89; <label>:31: ; preds = %30, %2990 %32 = call signext i32 undef(ptr undef, i64 %16, i32 signext 8)91 %33 = icmp eq i64 undef, 092 %34 = select i1 %33, i64 0, i64 %1993 %35 = add i64 %34, %1694 br i1 %33, label %36, label %1495 96; <label>:36: ; preds = %31, %197 ret void98}99 100declare fp128 @llvm.pow.f128(fp128, fp128)101 102define void @fun1(ptr) {103; CHECK-LABEL: .LBB1_2104; CHECK: =>This Inner Loop Header: Depth=1105; CHECK-NOT: 16-byte Folded Spill106; CHECK-NOT: 16-byte Folded Reload107; CHECK-LABEL: .LBB1_3108 109 br i1 undef, label %7, label %2110 111; <label>:2: ; preds = %2, %1112 %3 = phi fp128 [ %5, %2 ], [ 0xL00000000000000000000000000000000, %1 ]113 %4 = tail call fp128 @llvm.pow.f128(fp128 0xL00000000000000000000000000000000, fp128 0xL00000000000000000000000000000000) #2114 %5 = fadd fp128 %3, %4115 %6 = icmp eq i64 undef, 0116 br i1 %6, label %7, label %2117 118; <label>:7: ; preds = %2, %1119 %8 = phi fp128 [ 0xL00000000000000000000000000000000, %1 ], [ %5, %2 ]120 %9 = fadd fp128 0xL00000000000000000000000000000000, %8121 %10 = fadd fp128 0xL00000000000000000000000000000000, %9122 %11 = fadd fp128 0xL00000000000000000000000000000000, %10123 %12 = tail call fp128 @llvm.pow.f128(fp128 %11, fp128 0xL00000000000000000000000000000000) #2124 store fp128 %12, ptr %0, align 8125 ret void126}127