74 lines · plain
1; RUN: llc -mtriple=amdgcn -mcpu=bonaire -show-mc-encoding < %s | FileCheck --check-prefixes=GCN,CI,ALL %s2; RUN: llc -mtriple=amdgcn -mcpu=carrizo --show-mc-encoding < %s | FileCheck --check-prefixes=GCN,VI,ALL %s3; RUN: llc -mtriple=amdgcn -mcpu=gfx900 --show-mc-encoding < %s | FileCheck --check-prefixes=GCN,GFX9,ALL %s4; RUN: llc -mcpu=bonaire -mtriple=amdgcn-unknown-amdhsa -mattr=-flat-for-global < %s | FileCheck --check-prefixes=GCNHSA,ALL %s5; RUN: llc -mcpu=carrizo -mtriple=amdgcn-unknown-amdhsa -mattr=-flat-for-global < %s | FileCheck --check-prefixes=GCNHSA,ALL %s6; RUN: llc -mcpu=gfx1010 -mtriple=amdgcn-unknown-amdhsa -mattr=-flat-for-global < %s | FileCheck --check-prefixes=GCNHSA,ALL %s7; RUN: llc -mcpu=gfx1100 -mtriple=amdgcn-unknown-amdhsa -mattr=-flat-for-global,-architected-flat-scratch,-user-sgpr-init16-bug < %s | FileCheck --check-prefixes=GCNHSA,ALL %s8 9; FIXME: align on alloca seems to be ignored for private_segment_alignment10 11; ALL-LABEL: {{^}}large_alloca_compute_shader:12 13; GCN-DAG: s_mov_b32 s{{[0-9]+}}, SCRATCH_RSRC_DWORD014; GCN-DAG: ; fixup A - offset: 4, value: SCRATCH_RSRC_DWORD015; GCN-DAG: s_mov_b32 s{{[0-9]+}}, SCRATCH_RSRC_DWORD116; GCN-DAG: ; fixup A - offset: 4, value: SCRATCH_RSRC_DWORD117; GCN-DAG: s_mov_b32 s{{[0-9]+}}, -118; CI-DAG: s_mov_b32 s{{[0-9]+}}, 0xe8f00019; VI-DAG: s_mov_b32 s{{[0-9]+}}, 0xe8000020; GFX9-DAG: s_mov_b32 s{{[0-9]+}}, 0xe0000021 22; GCNHSA: buffer_store_{{dword|b32}} {{v[0-9]+}}, {{v[0-9]+}}, s[0:3], 0 offen23; GCNHSA: buffer_load_{{dword|b32}} {{v[0-9]+}}, {{v[0-9]+}}, s[0:3], 0 offen24 25; GCNHSA: .amdhsa_kernel large_alloca_compute_shader26; GCNHSA: .amdhsa_group_segment_fixed_size 027; GCNHSA: .amdhsa_private_segment_fixed_size 3277228; GCNHSA: .amdhsa_user_sgpr_private_segment_buffer 129; GCNHSA: .amdhsa_user_sgpr_dispatch_ptr 130; GCNHSA: .amdhsa_user_sgpr_queue_ptr 131; GCNHSA: .amdhsa_user_sgpr_kernarg_segment_ptr 132; GCNHSA: .amdhsa_user_sgpr_dispatch_id 133; GCNHSA: .amdhsa_user_sgpr_flat_scratch_init 134; GCNHSA: .amdhsa_user_sgpr_private_segment_size 035; GCNHSA: .amdhsa_system_sgpr_private_segment_wavefront_offset 136; GCNHSA: .amdhsa_system_sgpr_workgroup_id_x 137; GCNHSA: .amdhsa_system_sgpr_workgroup_id_y 138; GCNHSA: .amdhsa_system_sgpr_workgroup_id_z 139; GCNHSA: .amdhsa_system_sgpr_workgroup_info 040; GCNHSA: .amdhsa_system_vgpr_workitem_id 241; GCNHSA: .amdhsa_next_free_vgpr {{2|3}}42; GCNHSA: .amdhsa_next_free_sgpr 1843; GCNHSA: .amdhsa_float_round_mode_32 044; GCNHSA: .amdhsa_float_round_mode_16_64 045; GCNHSA: .amdhsa_float_denorm_mode_32 346; GCNHSA: .amdhsa_float_denorm_mode_16_64 347; GCNHSA: .amdhsa_dx10_clamp 148; GCNHSA: .amdhsa_ieee_mode 149; GCNHSA: .amdhsa_exception_fp_ieee_invalid_op 050; GCNHSA: .amdhsa_exception_fp_denorm_src 051; GCNHSA: .amdhsa_exception_fp_ieee_div_zero 052; GCNHSA: .amdhsa_exception_fp_ieee_overflow 053; GCNHSA: .amdhsa_exception_fp_ieee_underflow 054; GCNHSA: .amdhsa_exception_fp_ieee_inexact 055; GCNHSA: .amdhsa_exception_int_div_zero 056; GCNHSA: .end_amdhsa_kernel57 58; Scratch size = alloca size + emergency stack slot, align {{.*}}, addrspace(5)59; ALL: ; ScratchSize: 3277260define amdgpu_kernel void @large_alloca_compute_shader(i32 %x, i32 %y) #0 {61 %large = alloca [8192 x i32], align 4, addrspace(5)62 %gep = getelementptr [8192 x i32], ptr addrspace(5) %large, i32 0, i32 819163 store volatile i32 %x, ptr addrspace(5) %gep64 %gep1 = getelementptr [8192 x i32], ptr addrspace(5) %large, i32 0, i32 %y65 %val = load volatile i32, ptr addrspace(5) %gep166 store volatile i32 %val, ptr addrspace(1) poison67 ret void68}69 70attributes #0 = { nounwind }71 72!llvm.module.flags = !{!0}73!0 = !{i32 1, !"amdhsa_code_object_version", i32 400}74