1 ; NOTE: Assertions have been autogenerated by utils/update_llc_test_checks.py
2 ; RUN: opt -S -passes='require<profile-summary>,function(codegenprepare)' -mtriple=amdgcn-amd-amdhsa -mcpu=gfx908 < %s | FileCheck -check-prefix=OPT %s
3 ; RUN: llc -mtriple=amdgcn-amd-amdhsa -mcpu=gfx908 < %s | FileCheck -check-prefix=GCN %s
5 ; Make sure we match the addressing mode offset of globla.atomic.fadd intrinsics across blocks.
7 define amdgpu_kernel void @test_sink_small_offset_global_atomic_fadd_f32(ptr addrspace(1) %out, ptr addrspace(1) %in) {
8 ; OPT-LABEL: @test_sink_small_offset_global_atomic_fadd_f32(
10 ; OPT-NEXT: [[TID:%.*]] = call i32 @llvm.amdgcn.mbcnt.lo(i32 -1, i32 0) #[[ATTR3:[0-9]+]]
11 ; OPT-NEXT: [[CMP:%.*]] = icmp eq i32 [[TID]], 0
12 ; OPT-NEXT: br i1 [[CMP]], label [[ENDIF:%.*]], label [[IF:%.*]]
14 ; OPT-NEXT: [[IN_GEP:%.*]] = getelementptr float, ptr addrspace(1) [[IN:%.*]], i32 7
15 ; OPT-NEXT: [[FADD2:%.*]] = call float @llvm.amdgcn.global.atomic.fadd.f32.p1.f32(ptr addrspace(1) [[IN_GEP]], float 2.000000e+00)
16 ; OPT-NEXT: [[VAL:%.*]] = load volatile float, ptr addrspace(1) undef, align 4
17 ; OPT-NEXT: br label [[ENDIF]]
19 ; OPT-NEXT: [[X:%.*]] = phi float [ [[VAL]], [[IF]] ], [ 0.000000e+00, [[ENTRY:%.*]] ]
20 ; OPT-NEXT: [[OUT_GEP:%.*]] = getelementptr float, ptr addrspace(1) [[OUT:%.*]], i32 999999
21 ; OPT-NEXT: store float [[X]], ptr addrspace(1) [[OUT_GEP]], align 4
22 ; OPT-NEXT: br label [[DONE:%.*]]
26 ; GCN-LABEL: test_sink_small_offset_global_atomic_fadd_f32:
27 ; GCN: ; %bb.0: ; %entry
28 ; GCN-NEXT: s_load_dwordx4 s[0:3], s[6:7], 0x0
29 ; GCN-NEXT: v_mbcnt_lo_u32_b32 v0, -1, 0
30 ; GCN-NEXT: v_cmp_ne_u32_e32 vcc, 0, v0
31 ; GCN-NEXT: v_mov_b32_e32 v0, 0
32 ; GCN-NEXT: s_and_saveexec_b64 s[4:5], vcc
33 ; GCN-NEXT: s_cbranch_execz .LBB0_2
34 ; GCN-NEXT: ; %bb.1: ; %if
35 ; GCN-NEXT: v_mov_b32_e32 v0, 0
36 ; GCN-NEXT: v_mov_b32_e32 v1, 2.0
37 ; GCN-NEXT: s_waitcnt lgkmcnt(0)
38 ; GCN-NEXT: global_atomic_add_f32 v0, v1, s[2:3] offset:28
39 ; GCN-NEXT: global_load_dword v0, v[0:1], off glc
40 ; GCN-NEXT: s_waitcnt vmcnt(0)
41 ; GCN-NEXT: .LBB0_2: ; %endif
42 ; GCN-NEXT: s_or_b64 exec, exec, s[4:5]
43 ; GCN-NEXT: v_mov_b32_e32 v1, 0x3d0000
44 ; GCN-NEXT: s_waitcnt lgkmcnt(0)
45 ; GCN-NEXT: global_store_dword v1, v0, s[0:1] offset:2300
48 %tid = call i32 @llvm.amdgcn.mbcnt.lo(i32 -1, i32 0) #0
49 %cmp = icmp eq i32 %tid, 0
50 br i1 %cmp, label %endif, label %if
53 %in.gep = getelementptr float, ptr addrspace(1) %in, i32 7
54 %fadd2 = call float @llvm.amdgcn.global.atomic.fadd.f32.p1.f32(ptr addrspace(1) %in.gep, float 2.0)
55 %val = load volatile float, ptr addrspace(1) undef
59 %x = phi float [ %val, %if ], [ 0.0, %entry ]
60 %out.gep = getelementptr float, ptr addrspace(1) %out, i32 999999
61 store float %x, ptr addrspace(1) %out.gep
68 declare i32 @llvm.amdgcn.mbcnt.lo(i32, i32) #1
69 declare float @llvm.amdgcn.global.atomic.fadd.f32.p1.f32(ptr addrspace(1) nocapture, float) #2
71 attributes #0 = { argmemonly nounwind }
72 attributes #1 = { nounwind readnone willreturn }
73 attributes #2 = { argmemonly nounwind willreturn }