125 lines · plain
1; RUN: llc -mtriple=amdgcn -mattr=-promote-alloca,+max-private-element-size-16 < %s | FileCheck --check-prefixes=SI-ALLOCA16,SI %s2; RUN: llc -mtriple=amdgcn -mattr=-promote-alloca,+max-private-element-size-4 < %s | FileCheck --check-prefixes=SI-ALLOCA4,SI %s3; RUN: llc -mtriple=amdgcn -mattr=+promote-alloca -disable-promote-alloca-to-vector < %s | FileCheck --check-prefixes=SI-PROMOTE,SI %s4; RUN: llc -mtriple=amdgcn -mcpu=tonga -mattr=-flat-for-global -mattr=-promote-alloca,+max-private-element-size-16 < %s | FileCheck -check-prefix=SI %s5; RUN: llc -mtriple=amdgcn -mcpu=tonga -mattr=-flat-for-global -mattr=+promote-alloca -disable-promote-alloca-to-vector < %s | FileCheck --check-prefixes=CI-PROMOTE,SI %s6 7declare void @llvm.amdgcn.s.barrier() #08 9; SI-LABEL: {{^}}private_access_f64_alloca:10 11; SI-ALLOCA16: buffer_store_dwordx212; SI-ALLOCA16: buffer_load_dwordx213 14; SI-ALLOCA4: buffer_store_dword v15; SI-ALLOCA4: buffer_store_dword v16; SI-ALLOCA4: buffer_load_dword v17; SI-ALLOCA4: buffer_load_dword v18 19; SI-PROMOTE: ds_write_b6420; SI-PROMOTE: ds_read_b6421; CI-PROMOTE: ds_write_b6422; CI-PROMOTE: ds_read_b6423define amdgpu_kernel void @private_access_f64_alloca(ptr addrspace(1) noalias %out, ptr addrspace(1) noalias %in, i32 %b) #1 {24 %val = load double, ptr addrspace(1) %in, align 825 %array = alloca [8 x double], align 8, addrspace(5)26 %ptr = getelementptr inbounds [8 x double], ptr addrspace(5) %array, i32 0, i32 %b27 store double %val, ptr addrspace(5) %ptr, align 828 call void @llvm.amdgcn.s.barrier()29 %result = load double, ptr addrspace(5) %ptr, align 830 store double %result, ptr addrspace(1) %out, align 831 ret void32}33 34; SI-LABEL: {{^}}private_access_v2f64_alloca:35 36; SI-ALLOCA16: buffer_store_dwordx437; SI-ALLOCA16: buffer_load_dwordx438 39; SI-ALLOCA4: buffer_store_dword v40; SI-ALLOCA4: buffer_store_dword v41; SI-ALLOCA4: buffer_store_dword v42; SI-ALLOCA4: buffer_store_dword v43; SI-ALLOCA4: buffer_load_dword v44; SI-ALLOCA4: buffer_load_dword v45; SI-ALLOCA4: buffer_load_dword v46; SI-ALLOCA4: buffer_load_dword v47 48; SI-PROMOTE: ds_write_b6449; SI-PROMOTE: ds_write_b6450; SI-PROMOTE: ds_read_b6451; SI-PROMOTE: ds_read_b6452; CI-PROMOTE: ds_write_b12853; CI-PROMOTE: ds_read_b12854define amdgpu_kernel void @private_access_v2f64_alloca(ptr addrspace(1) noalias %out, ptr addrspace(1) noalias %in, i32 %b) #1 {55 %val = load <2 x double>, ptr addrspace(1) %in, align 1656 %array = alloca [4 x <2 x double>], align 16, addrspace(5)57 %ptr = getelementptr inbounds [4 x <2 x double>], ptr addrspace(5) %array, i32 0, i32 %b58 store <2 x double> %val, ptr addrspace(5) %ptr, align 1659 call void @llvm.amdgcn.s.barrier()60 %result = load <2 x double>, ptr addrspace(5) %ptr, align 1661 store <2 x double> %result, ptr addrspace(1) %out, align 1662 ret void63}64 65; SI-LABEL: {{^}}private_access_i64_alloca:66 67; SI-ALLOCA16: buffer_store_dwordx268; SI-ALLOCA16: buffer_load_dwordx269 70; SI-ALLOCA4: buffer_store_dword v71; SI-ALLOCA4: buffer_store_dword v72; SI-ALLOCA4: buffer_load_dword v73; SI-ALLOCA4: buffer_load_dword v74 75 76; SI-PROMOTE: ds_write_b6477; SI-PROMOTE: ds_read_b6478; CI-PROMOTE: ds_write_b6479; CI-PROMOTE: ds_read_b6480define amdgpu_kernel void @private_access_i64_alloca(ptr addrspace(1) noalias %out, ptr addrspace(1) noalias %in, i32 %b) #1 {81 %val = load i64, ptr addrspace(1) %in, align 882 %array = alloca [8 x i64], align 8, addrspace(5)83 %ptr = getelementptr inbounds [8 x i64], ptr addrspace(5) %array, i32 0, i32 %b84 store i64 %val, ptr addrspace(5) %ptr, align 885 call void @llvm.amdgcn.s.barrier()86 %result = load i64, ptr addrspace(5) %ptr, align 887 store i64 %result, ptr addrspace(1) %out, align 888 ret void89}90 91; SI-LABEL: {{^}}private_access_v2i64_alloca:92 93; SI-ALLOCA16: buffer_store_dwordx494; SI-ALLOCA16: buffer_load_dwordx495 96; SI-ALLOCA4: buffer_store_dword v97; SI-ALLOCA4: buffer_store_dword v98; SI-ALLOCA4: buffer_store_dword v99; SI-ALLOCA4: buffer_store_dword v100 101; SI-ALLOCA4: buffer_load_dword v102; SI-ALLOCA4: buffer_load_dword v103; SI-ALLOCA4: buffer_load_dword v104; SI-ALLOCA4: buffer_load_dword v105 106; SI-PROMOTE: ds_write_b64107; SI-PROMOTE: ds_write_b64108; SI-PROMOTE: ds_read_b64109; SI-PROMOTE: ds_read_b64110; CI-PROMOTE: ds_write_b128111; CI-PROMOTE: ds_read_b128112define amdgpu_kernel void @private_access_v2i64_alloca(ptr addrspace(1) noalias %out, ptr addrspace(1) noalias %in, i32 %b) #1 {113 %val = load <2 x i64>, ptr addrspace(1) %in, align 16114 %array = alloca [4 x <2 x i64>], align 16, addrspace(5)115 %ptr = getelementptr inbounds [4 x <2 x i64>], ptr addrspace(5) %array, i32 0, i32 %b116 store <2 x i64> %val, ptr addrspace(5) %ptr, align 16117 call void @llvm.amdgcn.s.barrier()118 %result = load <2 x i64>, ptr addrspace(5) %ptr, align 16119 store <2 x i64> %result, ptr addrspace(1) %out, align 16120 ret void121}122 123attributes #0 = { convergent nounwind }124attributes #1 = { nounwind "amdgpu-waves-per-eu"="1,2" "amdgpu-flat-work-group-size"="64,128" }125