1; NOTE: Assertions have been autogenerated by utils/update_test_checks.py
2; NOTE: Assertions have been autogenerated by utils/update_llc_test_checks.py
3; RUN: opt -S -codegenprepare -mtriple=amdgcn-amd-amdhsa -mcpu=gfx1030 < %s | FileCheck -check-prefix=OPT %s
4; RUN: llc -mtriple=amdgcn-amd-amdhsa -mcpu=gfx1030 < %s | FileCheck -check-prefix=GCN %s
5
6; Make sure we match the addressing mode offset of csub intrinsics across blocks.
7
8define amdgpu_kernel void @test_sink_small_offset_global_atomic_csub_i32(i32 addrspace(1)* %out, i32 addrspace(1)* %in) {
9; OPT-LABEL: @test_sink_small_offset_global_atomic_csub_i32(
10; OPT-NEXT:  entry:
11; OPT-NEXT:    [[OUT_GEP:%.*]] = getelementptr i32, i32 addrspace(1)* [[OUT:%.*]], i32 999999
12; OPT-NEXT:    [[TID:%.*]] = call i32 @llvm.amdgcn.mbcnt.lo(i32 -1, i32 0) #3
13; OPT-NEXT:    [[CMP:%.*]] = icmp eq i32 [[TID]], 0
14; OPT-NEXT:    br i1 [[CMP]], label [[ENDIF:%.*]], label [[IF:%.*]]
15; OPT:       if:
16; OPT-NEXT:    [[TMP0:%.*]] = bitcast i32 addrspace(1)* [[IN:%.*]] to i8 addrspace(1)*
17; OPT-NEXT:    [[SUNKADDR:%.*]] = getelementptr i8, i8 addrspace(1)* [[TMP0]], i64 28
18; OPT-NEXT:    [[TMP1:%.*]] = bitcast i8 addrspace(1)* [[SUNKADDR]] to i32 addrspace(1)*
19; OPT-NEXT:    [[VAL:%.*]] = call i32 @llvm.amdgcn.global.atomic.csub.p1i32(i32 addrspace(1)* [[TMP1]], i32 2)
20; OPT-NEXT:    br label [[ENDIF]]
21; OPT:       endif:
22; OPT-NEXT:    [[X:%.*]] = phi i32 [ [[VAL]], [[IF]] ], [ 0, [[ENTRY:%.*]] ]
23; OPT-NEXT:    store i32 [[X]], i32 addrspace(1)* [[OUT_GEP]], align 4
24; OPT-NEXT:    ret void
25;
26; GCN-LABEL: test_sink_small_offset_global_atomic_csub_i32:
27; GCN:       ; %bb.0: ; %entry
28; GCN-NEXT:    s_load_dwordx4 s[0:3], s[4:5], 0x0
29; GCN-NEXT:    v_mbcnt_lo_u32_b32_e64 v0, -1, 0
30; GCN-NEXT:    v_cmp_ne_u32_e32 vcc_lo, 0, v0
31; GCN-NEXT:    v_mov_b32_e32 v0, 0
32; GCN-NEXT:    s_and_saveexec_b32 s4, vcc_lo
33; GCN-NEXT:    s_cbranch_execz BB0_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
37; GCN-NEXT:    s_waitcnt lgkmcnt(0)
38; GCN-NEXT:    global_atomic_csub v0, v0, v1, s[2:3] offset:28 glc
39; GCN-NEXT:  BB0_2: ; %endif
40; GCN-NEXT:    s_or_b32 exec_lo, exec_lo, s4
41; GCN-NEXT:    v_mov_b32_e32 v1, 0x3d0800
42; GCN-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
43; GCN-NEXT:    global_store_dword v1, v0, s[0:1] offset:252
44; GCN-NEXT:    s_endpgm
45entry:
46  %out.gep = getelementptr i32, i32 addrspace(1)* %out, i32 999999
47  %in.gep = getelementptr i32, i32 addrspace(1)* %in, i32 7
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
51
52if:
53  %val = call i32 @llvm.amdgcn.global.atomic.csub.p1i32(i32 addrspace(1)* %in.gep, i32 2)
54  br label %endif
55
56endif:
57  %x = phi i32 [ %val, %if ], [ 0, %entry ]
58  store i32 %x, i32 addrspace(1)* %out.gep
59  br label %done
60
61done:
62  ret void
63}
64
65declare i32 @llvm.amdgcn.global.atomic.csub.p1i32(i32 addrspace(1)* nocapture, i32) #0
66declare i32 @llvm.amdgcn.mbcnt.lo(i32, i32) #1
67
68attributes #0 = { argmemonly nounwind }
69attributes #1 = { nounwind readnone willreturn }
70attributes #2 = { argmemonly nounwind willreturn }
71