1 ; RUN: llc -march=amdgcn -mcpu=tonga -mattr=-flat-for-global -verify-machineinstrs -show-mc-encoding < %s | FileCheck -check-prefixes=VI,VI-OPT,PREGFX10,PREGFX10-OPT %s
2 ; RUN: llc -O0 -march=amdgcn -mcpu=tonga -mattr=-flat-for-global -verify-machineinstrs -show-mc-encoding < %s | FileCheck -check-prefixes=VI,VI-NOOPT,PREGFX10,PREGFX10-NOOPT %s
3 ; RUN: llc -march=amdgcn -mcpu=gfx1010 -mattr=-flat-for-global -verify-machineinstrs -show-mc-encoding < %s | FileCheck -check-prefixes=VI,VI-OPT %s
5 ; FIXME: The register allocator / scheduler should be able to avoid these hazards.
7 ; VI-LABEL: {{^}}dpp_test:
8 ; VI: v_mov_b32_e32 v0, s{{[0-9]+}}
9 ; VI-NOOPT: v_mov_b32_e32 v1, s{{[0-9]+}}
10 ; PREGFX10-OPT: s_nop 1
11 ; PREGFX10-NOOPT: s_nop 0
12 ; PREGFX10-NOOPT: s_nop 0
13 ; VI-OPT: 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]
14 ; VI-NOOPT: v_mov_b32_dpp v0, v1 quad_perm:[1,0,0,0] row_mask:0x1 bank_mask:0x1 bound_ctrl:0 ; encoding: [0xfa,0x02,0x00,0x7e,0x01,0x01,0x08,0x11]
15 define amdgpu_kernel void @dpp_test(i32 addrspace(1)* %out, i32 %in) {
16 %tmp0 = call i32 @llvm.amdgcn.mov.dpp.i32(i32 %in, i32 1, i32 1, i32 1, i1 1) #0
17 store i32 %tmp0, i32 addrspace(1)* %out
21 ; VI-LABEL: {{^}}dpp_wait_states:
22 ; VI-NOOPT: v_mov_b32_e32 [[VGPR1:v[0-9]+]], s{{[0-9]+}}
23 ; VI: v_mov_b32_e32 [[VGPR0:v[0-9]+]], s{{[0-9]+}}
24 ; PREGFX10-OPT: s_nop 1
25 ; PREGFX10-NOOPT: s_nop 0
26 ; PREGFX10-NOOPT: s_nop 0
27 ; VI-OPT: v_mov_b32_dpp [[VGPR0]], [[VGPR0]] quad_perm:[1,0,0,0] row_mask:0x1 bank_mask:0x1 bound_ctrl:0
28 ; VI-NOOPT: v_mov_b32_dpp [[VGPR1]], [[VGPR0]] quad_perm:[1,0,0,0] row_mask:0x1 bank_mask:0x1 bound_ctrl:
29 ; PREGFX10-OPT: s_nop 1
30 ; PREGFX10-NOOPT: s_nop 0
31 ; PREGFX10-NOOPT: s_nop 0
32 ; VI-OPT: v_mov_b32_dpp v{{[0-9]+}}, [[VGPR0]] quad_perm:[1,0,0,0] row_mask:0x1 bank_mask:0x1 bound_ctrl:0
33 ; VI-NOOPT: v_mov_b32_dpp v{{[0-9]+}}, [[VGPR1]] quad_perm:[1,0,0,0] row_mask:0x1 bank_mask:0x1 bound_ctrl:0
34 define amdgpu_kernel void @dpp_wait_states(i32 addrspace(1)* %out, i32 %in) {
35 %tmp0 = call i32 @llvm.amdgcn.mov.dpp.i32(i32 %in, i32 1, i32 1, i32 1, i1 1) #0
36 %tmp1 = call i32 @llvm.amdgcn.mov.dpp.i32(i32 %tmp0, i32 1, i32 1, i32 1, i1 1) #0
37 store i32 %tmp1, i32 addrspace(1)* %out
41 ; VI-LABEL: {{^}}dpp_first_in_bb:
43 ; PREGFX10-OPT: s_mov_b32
44 ; PREGFX10-OPT: s_mov_b32
45 ; PREGFX10-NOOPT: s_waitcnt
46 ; PREGFX10-NOOPT-NEXT: s_nop 0
47 ; 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
48 ; PREGFX10-OPT: s_nop 1
49 ; 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
50 ; PREGFX10-OPT: s_nop 1
51 ; PREGFX10-NOOPT: s_nop 0
52 ; PREGFX10-NOOPT: s_nop 0
53 ; VI: v_mov_b32_dpp v{{[0-9]+}}, [[VGPR1]] quad_perm:[1,0,0,0] row_mask:0x1 bank_mask:0x1 bound_ctrl:0
54 define amdgpu_kernel void @dpp_first_in_bb(float addrspace(1)* %out, float addrspace(1)* %in, float %cond, float %a, float %b) {
55 %cmp = fcmp oeq float %cond, 0.0
56 br i1 %cmp, label %if, label %else
59 %out_val = load float, float addrspace(1)* %out
60 %if_val = fadd float %a, %out_val
64 %in_val = load float, float addrspace(1)* %in
65 %else_val = fadd float %b, %in_val
69 %val = phi float [%if_val, %if], [%else_val, %else]
70 %val_i32 = bitcast float %val to i32
71 %tmp0 = call i32 @llvm.amdgcn.mov.dpp.i32(i32 %val_i32, i32 1, i32 1, i32 1, i1 1) #0
72 %tmp1 = call i32 @llvm.amdgcn.mov.dpp.i32(i32 %tmp0, i32 1, i32 1, i32 1, i1 1) #0
73 %tmp2 = call i32 @llvm.amdgcn.mov.dpp.i32(i32 %tmp1, i32 1, i32 1, i32 1, i1 1) #0
74 %tmp_float = bitcast i32 %tmp2 to float
75 store float %tmp_float, float addrspace(1)* %out
79 declare i32 @llvm.amdgcn.mov.dpp.i32(i32, i32, i32, i32, i1) #0
81 attributes #0 = { nounwind readnone convergent }