56 lines · plain
1# NOTE: Assertions have been autogenerated by utils/update_mir_test_checks.py2# RUN: llc -mtriple=amdgcn -mcpu=gfx900 -verify-machineinstrs -run-pass=si-insert-waitcnts %s -o - | FileCheck -check-prefix=GCN %s3# RUN: llc -mtriple=amdgcn -mcpu=gfx900 -passes=si-insert-waitcnts %s -o - | FileCheck -check-prefix=GCN %s4 5# $sgpr30_sgpr31 will hold the return address. We need a waitcnt before SI_CALL so6# that the return address is not clobbered in the callee by the outstanding load.7 8--- |9 define amdgpu_kernel void @call_waw_waitcnt() #0 {10 %1 = call i32 @func()11 ret void12 }13 14 declare hidden i32 @func() #015 16 attributes #0 = { nounwind }17...18 19---20name: call_waw_waitcnt21tracksRegLiveness: true22body: |23 bb.0:24 liveins: $sgpr4_sgpr5, $sgpr7, $sgpr0_sgpr1_sgpr2_sgpr325 26 ; GCN-LABEL: name: call_waw_waitcnt27 ; GCN: liveins: $sgpr4_sgpr5, $sgpr7, $sgpr0_sgpr1_sgpr2_sgpr328 ; GCN-NEXT: {{ $}}29 ; GCN-NEXT: S_WAITCNT 030 ; GCN-NEXT: $sgpr30_sgpr31 = S_LOAD_DWORDX2_IMM $sgpr4_sgpr5, 0, 031 ; GCN-NEXT: $sgpr33 = S_MOV_B32 killed $sgpr732 ; GCN-NEXT: $flat_scr_lo = S_ADD_U32 killed $sgpr4, $sgpr33, implicit-def $scc33 ; GCN-NEXT: $flat_scr_hi = S_ADDC_U32 killed $sgpr5, 0, implicit-def $scc, implicit killed $scc34 ; GCN-NEXT: BUNDLE implicit-def $sgpr4_sgpr5, implicit-def $sgpr4, implicit-def $sgpr5, implicit-def $scc {35 ; GCN-NEXT: $sgpr4_sgpr5 = S_GETPC_B6436 ; GCN-NEXT: $sgpr4 = S_ADD_U32 internal $sgpr4, target-flags(amdgpu-rel32-lo) @func + 4, implicit-def $scc37 ; GCN-NEXT: $sgpr5 = S_ADDC_U32 internal $sgpr5, target-flags(amdgpu-rel32-hi) @func + 4, implicit-def $scc, implicit internal $scc38 ; GCN-NEXT: }39 ; GCN-NEXT: $sgpr32 = S_MOV_B32 $sgpr3340 ; GCN-NEXT: S_WAITCNT 4927941 ; GCN-NEXT: dead $sgpr30_sgpr31 = SI_CALL killed renamable $sgpr4_sgpr5, @func, csr_amdgpu, implicit $sgpr0_sgpr1_sgpr2_sgpr3, implicit-def dead $vgpr042 ; GCN-NEXT: S_ENDPGM 043 $sgpr30_sgpr31 = S_LOAD_DWORDX2_IMM $sgpr4_sgpr5, 0, 044 $sgpr33 = S_MOV_B32 killed $sgpr745 $flat_scr_lo = S_ADD_U32 killed $sgpr4, $sgpr33, implicit-def $scc46 $flat_scr_hi = S_ADDC_U32 killed $sgpr5, 0, implicit-def $scc, implicit killed $scc47 BUNDLE implicit-def $sgpr4_sgpr5, implicit-def $sgpr4, implicit-def $sgpr5, implicit-def $scc {48 $sgpr4_sgpr5 = S_GETPC_B6449 $sgpr4 = S_ADD_U32 internal $sgpr4, target-flags(amdgpu-rel32-lo) @func + 4, implicit-def $scc50 $sgpr5 = S_ADDC_U32 internal $sgpr5, target-flags(amdgpu-rel32-hi) @func + 4, implicit-def $scc, implicit internal $scc51 }52 $sgpr32 = S_MOV_B32 $sgpr3353 dead $sgpr30_sgpr31 = SI_CALL killed renamable $sgpr4_sgpr5, @func, csr_amdgpu, implicit $sgpr0_sgpr1_sgpr2_sgpr3, implicit-def dead $vgpr054 S_ENDPGM 055...56