xref: /aosp_15_r20/external/llvm/test/CodeGen/AMDGPU/llvm.amdgcn.mov.dpp.ll (revision 9880d6810fe72a1726cb53787c6711e909410d58)
1; RUN: llc -march=amdgcn -mcpu=tonga -verify-machineinstrs -show-mc-encoding < %s | FileCheck -check-prefix=VI -check-prefix=VI-OPT %s
2; RUN: llc -O0 -march=amdgcn -mcpu=tonga -verify-machineinstrs -show-mc-encoding < %s | FileCheck -check-prefix=VI -check-prefix=VI-NOOPT %s
3
4; FIXME: The register allocator / scheduler should be able to avoid these hazards.
5
6; VI-LABEL: {{^}}dpp_test:
7; VI: v_mov_b32_e32 v0, s{{[0-9]+}}
8; VI: s_nop 1
9; VI: v_mov_b32_dpp v0, v0 quad_perm:[1,0,0,0] row_mask:0x1 bank_mask:0x1 bound_ctrl:0 ; encoding: [0xfa,0x02,0x00,0x7e,0x00,0x01,0x08,0x11]
10define void @dpp_test(i32 addrspace(1)* %out, i32 %in) {
11  %tmp0 = call i32 @llvm.amdgcn.mov.dpp.i32(i32 %in, i32 1, i32 1, i32 1, i1 1) #0
12  store i32 %tmp0, i32 addrspace(1)* %out
13  ret void
14}
15
16; VI-LABEL: {{^}}dpp_wait_states:
17; VI: v_mov_b32_e32 [[VGPR0:v[0-9]+]], s{{[0-9]+}}
18; VI: s_nop 1
19; VI: v_mov_b32_dpp [[VGPR1:v[0-9]+]], [[VGPR0]] quad_perm:[1,0,0,0] row_mask:0x1 bank_mask:0x1 bound_ctrl:0
20; VI: s_nop 1
21; VI: v_mov_b32_dpp v{{[0-9]+}}, [[VGPR1]] quad_perm:[1,0,0,0] row_mask:0x1 bank_mask:0x1 bound_ctrl:0
22define void @dpp_wait_states(i32 addrspace(1)* %out, i32 %in) {
23  %tmp0 = call i32 @llvm.amdgcn.mov.dpp.i32(i32 %in, i32 1, i32 1, i32 1, i1 1) #0
24  %tmp1 = call i32 @llvm.amdgcn.mov.dpp.i32(i32 %tmp0, i32 1, i32 1, i32 1, i1 1) #0
25  store i32 %tmp1, i32 addrspace(1)* %out
26  ret void
27}
28
29; VI-LABEL: {{^}}dpp_first_in_bb:
30; VI: ; %endif
31; VI-OPT: s_mov_b32
32; VI-OPT: s_mov_b32
33; VI-NOOPT: s_nop 1
34; VI: v_mov_b32_dpp [[VGPR0:v[0-9]+]], v{{[0-9]+}} quad_perm:[1,0,0,0] row_mask:0x1 bank_mask:0x1 bound_ctrl:0
35; VI: s_nop 1
36; VI: v_mov_b32_dpp [[VGPR1:v[0-9]+]], [[VGPR0]] quad_perm:[1,0,0,0] row_mask:0x1 bank_mask:0x1 bound_ctrl:0
37; VI: s_nop 1
38; VI: v_mov_b32_dpp v{{[0-9]+}}, [[VGPR1]] quad_perm:[1,0,0,0] row_mask:0x1 bank_mask:0x1 bound_ctrl:0
39define void @dpp_first_in_bb(float addrspace(1)* %out, float addrspace(1)* %in, float %cond, float %a, float %b) {
40  %cmp = fcmp oeq float %cond, 0.0
41  br i1 %cmp, label %if, label %else
42
43if:
44  %out_val = load float, float addrspace(1)* %out
45  %if_val = fadd float %a, %out_val
46  br label %endif
47
48else:
49  %in_val = load float, float addrspace(1)* %in
50  %else_val = fadd float %b, %in_val
51  br label %endif
52
53endif:
54  %val = phi float [%if_val, %if], [%else_val, %else]
55  %val_i32 = bitcast float %val to i32
56  %tmp0 = call i32 @llvm.amdgcn.mov.dpp.i32(i32 %val_i32, i32 1, i32 1, i32 1, i1 1) #0
57  %tmp1 = call i32 @llvm.amdgcn.mov.dpp.i32(i32 %tmp0, i32 1, i32 1, i32 1, i1 1) #0
58  %tmp2 = call i32 @llvm.amdgcn.mov.dpp.i32(i32 %tmp1, i32 1, i32 1, i32 1, i1 1) #0
59  %tmp_float = bitcast i32 %tmp2 to float
60  store float %tmp_float, float addrspace(1)* %out
61  ret void
62}
63
64declare i32 @llvm.amdgcn.mov.dpp.i32(i32, i32, i32, i32, i1) #0
65
66attributes #0 = { nounwind readnone convergent }
67