1; NOTE: Assertions have been autogenerated by utils/update_llc_test_checks.py 2; RUN: llc -mtriple=aarch64-linux-gnu -mattr=+sme -verify-machineinstrs < %s | FileCheck %s 3 4declare void @private_za_callee() 5 6; Ensure that we don't use tail call optimization when a lazy-save is required. 7define void @disable_tailcallopt() "aarch64_inout_za" nounwind { 8; CHECK-LABEL: disable_tailcallopt: 9; CHECK: // %bb.0: 10; CHECK-NEXT: stp x29, x30, [sp, #-16]! // 16-byte Folded Spill 11; CHECK-NEXT: mov x29, sp 12; CHECK-NEXT: sub sp, sp, #16 13; CHECK-NEXT: rdsvl x8, #1 14; CHECK-NEXT: mov x9, sp 15; CHECK-NEXT: msub x9, x8, x8, x9 16; CHECK-NEXT: mov sp, x9 17; CHECK-NEXT: stur x9, [x29, #-16] 18; CHECK-NEXT: sub x9, x29, #16 19; CHECK-NEXT: sturh wzr, [x29, #-6] 20; CHECK-NEXT: stur wzr, [x29, #-4] 21; CHECK-NEXT: sturh w8, [x29, #-8] 22; CHECK-NEXT: msr TPIDR2_EL0, x9 23; CHECK-NEXT: bl private_za_callee 24; CHECK-NEXT: smstart za 25; CHECK-NEXT: mrs x8, TPIDR2_EL0 26; CHECK-NEXT: sub x0, x29, #16 27; CHECK-NEXT: cbnz x8, .LBB0_2 28; CHECK-NEXT: // %bb.1: 29; CHECK-NEXT: bl __arm_tpidr2_restore 30; CHECK-NEXT: .LBB0_2: 31; CHECK-NEXT: msr TPIDR2_EL0, xzr 32; CHECK-NEXT: mov sp, x29 33; CHECK-NEXT: ldp x29, x30, [sp], #16 // 16-byte Folded Reload 34; CHECK-NEXT: ret 35 tail call void @private_za_callee() 36 ret void 37} 38 39; Ensure we set up and restore the lazy save correctly for instructions which are lowered to lib calls 40define fp128 @f128_call_za(fp128 %a, fp128 %b) "aarch64_inout_za" nounwind { 41; CHECK-LABEL: f128_call_za: 42; CHECK: // %bb.0: 43; CHECK-NEXT: stp x29, x30, [sp, #-16]! // 16-byte Folded Spill 44; CHECK-NEXT: mov x29, sp 45; CHECK-NEXT: sub sp, sp, #16 46; CHECK-NEXT: rdsvl x8, #1 47; CHECK-NEXT: mov x9, sp 48; CHECK-NEXT: msub x9, x8, x8, x9 49; CHECK-NEXT: mov sp, x9 50; CHECK-NEXT: stur x9, [x29, #-16] 51; CHECK-NEXT: sub x9, x29, #16 52; CHECK-NEXT: sturh wzr, [x29, #-6] 53; CHECK-NEXT: stur wzr, [x29, #-4] 54; CHECK-NEXT: sturh w8, [x29, #-8] 55; CHECK-NEXT: msr TPIDR2_EL0, x9 56; CHECK-NEXT: bl __addtf3 57; CHECK-NEXT: smstart za 58; CHECK-NEXT: mrs x8, TPIDR2_EL0 59; CHECK-NEXT: sub x0, x29, #16 60; CHECK-NEXT: cbnz x8, .LBB1_2 61; CHECK-NEXT: // %bb.1: 62; CHECK-NEXT: bl __arm_tpidr2_restore 63; CHECK-NEXT: .LBB1_2: 64; CHECK-NEXT: msr TPIDR2_EL0, xzr 65; CHECK-NEXT: mov sp, x29 66; CHECK-NEXT: ldp x29, x30, [sp], #16 // 16-byte Folded Reload 67; CHECK-NEXT: ret 68 %res = fadd fp128 %a, %b 69 ret fp128 %res 70} 71