49 lines · plain
1; NOTE: Assertions have been autogenerated by utils/update_llc_test_checks.py UTC_ARGS: --version 52; RUN: llc -march=amdgcn -mcpu=gfx1100 < %s | FileCheck --check-prefixes=DEFAULT %s3; RUN: llc -march=amdgcn -mcpu=gfx1100 -amdgpu-waitcnt-load-forcezero < %s | FileCheck --check-prefixes=LDZERO %s4 5define amdgpu_kernel void @copy(ptr addrspace(1) noalias nocapture readonly %src1, ptr addrspace(1) noalias nocapture readonly %src2, ptr addrspace(1) noalias nocapture writeonly %dst1, ptr addrspace(1) noalias nocapture writeonly %dst2) {6; DEFAULT-LABEL: copy:7; DEFAULT: ; %bb.0:8; DEFAULT-NEXT: s_load_b256 s[0:7], s[4:5], 0x249; DEFAULT-NEXT: v_and_b32_e32 v0, 0x3ff, v010; DEFAULT-NEXT: s_delay_alu instid0(VALU_DEP_1)11; DEFAULT-NEXT: v_lshlrev_b32_e32 v0, 2, v012; DEFAULT-NEXT: s_waitcnt lgkmcnt(0)13; DEFAULT-NEXT: s_clause 0x114; DEFAULT-NEXT: global_load_b32 v1, v0, s[0:1]15; DEFAULT-NEXT: global_load_b32 v2, v0, s[2:3]16; DEFAULT-NEXT: s_waitcnt vmcnt(1)17; DEFAULT-NEXT: global_store_b32 v0, v1, s[4:5]18; DEFAULT-NEXT: s_waitcnt vmcnt(0)19; DEFAULT-NEXT: global_store_b32 v0, v2, s[6:7]20; DEFAULT-NEXT: s_endpgm21;22; LDZERO-LABEL: copy:23; LDZERO: ; %bb.0:24; LDZERO-NEXT: s_load_b256 s[0:7], s[4:5], 0x2425; LDZERO-NEXT: v_and_b32_e32 v0, 0x3ff, v026; LDZERO-NEXT: s_delay_alu instid0(VALU_DEP_1)27; LDZERO-NEXT: v_lshlrev_b32_e32 v0, 2, v028; LDZERO-NEXT: s_waitcnt lgkmcnt(0)29; LDZERO-NEXT: s_clause 0x130; LDZERO-NEXT: global_load_b32 v1, v0, s[0:1]31; LDZERO-NEXT: global_load_b32 v2, v0, s[2:3]32; LDZERO-NEXT: s_waitcnt vmcnt(0)33; LDZERO-NEXT: s_clause 0x134; LDZERO-NEXT: global_store_b32 v0, v1, s[4:5]35; LDZERO-NEXT: global_store_b32 v0, v2, s[6:7]36; LDZERO-NEXT: s_endpgm37 %id = tail call i32 @llvm.amdgcn.workitem.id.x()38 %idx = zext i32 %id to i6439 %gep.ld1 = getelementptr inbounds nuw float, ptr addrspace(1) %src1, i64 %idx40 %v1 = load float, ptr addrspace(1) %gep.ld1, align 441 %gep.ld2 = getelementptr inbounds nuw float, ptr addrspace(1) %src2, i64 %idx42 %v2 = load float, ptr addrspace(1) %gep.ld2, align 443 %gep.st1 = getelementptr inbounds nuw float, ptr addrspace(1) %dst1, i64 %idx44 store float %v1, ptr addrspace(1) %gep.st1, align 445 %gep.st2 = getelementptr inbounds nuw float, ptr addrspace(1) %dst2, i64 %idx46 store float %v2, ptr addrspace(1) %gep.st2, align 447 ret void48}49