1; RUN: llc -march=amdgcn -mtriple=amdgcn---amdgiz -amdgpu-atomic-optimizations=true -verify-machineinstrs < %s | FileCheck -enable-var-scope -check-prefixes=GCN,GCN64,GFX7LESS %s
2; RUN: llc -march=amdgcn -mtriple=amdgcn---amdgiz -mcpu=tonga -mattr=-flat-for-global -amdgpu-atomic-optimizations=true -verify-machineinstrs < %s | FileCheck -enable-var-scope -check-prefixes=GCN,GCN64,GFX8MORE,GFX8MORE64,GFX89,DPPCOMB %s
3; RUN: llc -march=amdgcn -mtriple=amdgcn---amdgiz -mcpu=gfx900 -mattr=-flat-for-global -amdgpu-atomic-optimizations=true -verify-machineinstrs < %s | FileCheck -enable-var-scope -check-prefixes=GCN,GCN64,GFX8MORE,GFX8MORE64,GFX89,DPPCOMB %s
4; RUN: llc -march=amdgcn -mtriple=amdgcn---amdgiz -mcpu=gfx1010 -mattr=-wavefrontsize32,+wavefrontsize64 -mattr=-flat-for-global -amdgpu-atomic-optimizations=true -verify-machineinstrs < %s | FileCheck -enable-var-scope -check-prefixes=GCN,GCN64,GFX8MORE,GFX8MORE64,GFX10 %s
5; RUN: llc -march=amdgcn -mtriple=amdgcn---amdgiz -mcpu=gfx1010 -mattr=+wavefrontsize32,-wavefrontsize64 -mattr=-flat-for-global -amdgpu-atomic-optimizations=true -verify-machineinstrs < %s | FileCheck -enable-var-scope -check-prefixes=GCN,GCN32,GFX8MORE,GFX8MORE32,GFX10 %s
6
7declare i32 @llvm.amdgcn.workitem.id.x()
8declare i32 @llvm.amdgcn.raw.buffer.atomic.add(i32, <4 x i32>, i32, i32, i32)
9declare i32 @llvm.amdgcn.raw.buffer.atomic.sub(i32, <4 x i32>, i32, i32, i32)
10
11; Show what the atomic optimization pass will do for raw buffers.
12
13; GCN-LABEL: add_i32_constant:
14; GCN32: s_mov_b32 s[[exec_lo:[0-9]+]], exec_lo
15; GCN64: s_mov_b64 s{{\[}}[[exec_lo:[0-9]+]]:[[exec_hi:[0-9]+]]{{\]}}, exec
16; GCN: v_mbcnt_lo_u32_b32{{(_e[0-9]+)?}} v[[mbcnt:[0-9]+]], s[[exec_lo]], 0
17; GCN64: v_mbcnt_hi_u32_b32{{(_e[0-9]+)?}} v[[mbcnt]], s[[exec_hi]], v[[mbcnt]]
18; GCN: v_cmp_eq_u32{{(_e[0-9]+)?}} vcc{{(_lo)?}}, 0, v[[mbcnt]]
19; GCN: s_and_saveexec_b{{32|64}} s[[exec:\[?[0-9:]+\]?]], vcc
20; GCN32: s_bcnt1_i32_b32 s[[popcount:[0-9]+]], s[[exec_lo]]
21; GCN64: s_bcnt1_i32_b64 s[[popcount:[0-9]+]], s{{\[}}[[exec_lo]]:[[exec_hi]]{{\]}}
22; GCN: v_mul_u32_u24{{(_e[0-9]+)?}} v[[value:[0-9]+]], s[[popcount]], 5
23; GCN: buffer_atomic_add v[[value]]
24define amdgpu_kernel void @add_i32_constant(i32 addrspace(1)* %out, <4 x i32> %inout) {
25entry:
26  %old = call i32 @llvm.amdgcn.raw.buffer.atomic.add(i32 5, <4 x i32> %inout, i32 0, i32 0, i32 0)
27  store i32 %old, i32 addrspace(1)* %out
28  ret void
29}
30
31; GCN-LABEL: add_i32_uniform:
32; GCN32: s_mov_b32 s[[exec_lo:[0-9]+]], exec_lo
33; GCN64: s_mov_b64 s{{\[}}[[exec_lo:[0-9]+]]:[[exec_hi:[0-9]+]]{{\]}}, exec
34; GCN: v_mbcnt_lo_u32_b32{{(_e[0-9]+)?}} v[[mbcnt:[0-9]+]], s[[exec_lo]], 0
35; GCN64: v_mbcnt_hi_u32_b32{{(_e[0-9]+)?}} v[[mbcnt]], s[[exec_hi]], v[[mbcnt]]
36; GCN: v_cmp_eq_u32{{(_e[0-9]+)?}} vcc{{(_lo)?}}, 0, v[[mbcnt]]
37; GCN: s_and_saveexec_b{{32|64}} s[[exec:\[?[0-9:]+\]?]], vcc
38; GCN32: s_bcnt1_i32_b32 s[[popcount:[0-9]+]], s[[exec_lo]]
39; GCN64: s_bcnt1_i32_b64 s[[popcount:[0-9]+]], s{{\[}}[[exec_lo]]:[[exec_hi]]{{\]}}
40; GCN: s_mul_i32 s[[scalar_value:[0-9]+]], s{{[0-9]+}}, s[[popcount]]
41; GCN: v_mov_b32{{(_e[0-9]+)?}} v[[value:[0-9]+]], s[[scalar_value]]
42; GCN: buffer_atomic_add v[[value]]
43define amdgpu_kernel void @add_i32_uniform(i32 addrspace(1)* %out, <4 x i32> %inout, i32 %additive) {
44entry:
45  %old = call i32 @llvm.amdgcn.raw.buffer.atomic.add(i32 %additive, <4 x i32> %inout, i32 0, i32 0, i32 0)
46  store i32 %old, i32 addrspace(1)* %out
47  ret void
48}
49
50; GCN-LABEL: add_i32_varying_vdata:
51; GFX7LESS-NOT: v_mbcnt_lo_u32_b32
52; GFX7LESS-NOT: v_mbcnt_hi_u32_b32
53; GFX7LESS-NOT: s_bcnt1_i32_b64
54; GFX7LESS: buffer_atomic_add v{{[0-9]+}}
55; DPPCOMB: v_add_u32_dpp
56; DPPCOMB: v_add_u32_dpp
57; GFX8MORE32: v_readlane_b32 s[[scalar_value:[0-9]+]], v{{[0-9]+}}, 31
58; GFX8MORE64: v_readlane_b32 s[[scalar_value:[0-9]+]], v{{[0-9]+}}, 63
59; GFX89: v_mov_b32{{(_e[0-9]+)?}} v[[value:[0-9]+]], s[[scalar_value]]
60; GFX10: s_mov_b32 s[[copy_value:[0-9]+]], s[[scalar_value]]
61; GFX10: v_mov_b32{{(_e[0-9]+)?}} v[[value:[0-9]+]], s[[copy_value]]
62; GFX8MORE: buffer_atomic_add v[[value]]
63define amdgpu_kernel void @add_i32_varying_vdata(i32 addrspace(1)* %out, <4 x i32> %inout) {
64entry:
65  %lane = call i32 @llvm.amdgcn.workitem.id.x()
66  %old = call i32 @llvm.amdgcn.raw.buffer.atomic.add(i32 %lane, <4 x i32> %inout, i32 0, i32 0, i32 0)
67  store i32 %old, i32 addrspace(1)* %out
68  ret void
69}
70
71; GCN-LABEL: add_i32_varying_offset:
72; GCN-NOT: v_mbcnt_lo_u32_b32
73; GCN-NOT: v_mbcnt_hi_u32_b32
74; GCN-NOT: s_bcnt1_i32_b64
75; GCN: buffer_atomic_add v{{[0-9]+}}
76define amdgpu_kernel void @add_i32_varying_offset(i32 addrspace(1)* %out, <4 x i32> %inout) {
77entry:
78  %lane = call i32 @llvm.amdgcn.workitem.id.x()
79  %old = call i32 @llvm.amdgcn.raw.buffer.atomic.add(i32 1, <4 x i32> %inout, i32 %lane, i32 0, i32 0)
80  store i32 %old, i32 addrspace(1)* %out
81  ret void
82}
83
84; GCN-LABEL: sub_i32_constant:
85; GCN32: s_mov_b32 s[[exec_lo:[0-9]+]], exec_lo
86; GCN64: s_mov_b64 s{{\[}}[[exec_lo:[0-9]+]]:[[exec_hi:[0-9]+]]{{\]}}, exec
87; GCN: v_mbcnt_lo_u32_b32{{(_e[0-9]+)?}} v[[mbcnt:[0-9]+]], s[[exec_lo]], 0
88; GCN64: v_mbcnt_hi_u32_b32{{(_e[0-9]+)?}} v[[mbcnt]], s[[exec_hi]], v[[mbcnt]]
89; GCN: v_cmp_eq_u32{{(_e[0-9]+)?}} vcc{{(_lo)?}}, 0, v[[mbcnt]]
90; GCN: s_and_saveexec_b{{32|64}} s[[exec:\[?[0-9:]+\]?]], vcc
91; GCN32: s_bcnt1_i32_b32 s[[popcount:[0-9]+]], s[[exec_lo]]
92; GCN64: s_bcnt1_i32_b64 s[[popcount:[0-9]+]], s{{\[}}[[exec_lo]]:[[exec_hi]]{{\]}}
93; GCN: v_mul_u32_u24{{(_e[0-9]+)?}} v[[value:[0-9]+]], s[[popcount]], 5
94; GCN: buffer_atomic_sub v[[value]]
95define amdgpu_kernel void @sub_i32_constant(i32 addrspace(1)* %out, <4 x i32> %inout) {
96entry:
97  %old = call i32 @llvm.amdgcn.raw.buffer.atomic.sub(i32 5, <4 x i32> %inout, i32 0, i32 0, i32 0)
98  store i32 %old, i32 addrspace(1)* %out
99  ret void
100}
101
102; GCN-LABEL: sub_i32_uniform:
103; GCN32: s_mov_b32 s[[exec_lo:[0-9]+]], exec_lo
104; GCN64: s_mov_b64 s{{\[}}[[exec_lo:[0-9]+]]:[[exec_hi:[0-9]+]]{{\]}}, exec
105; GCN: v_mbcnt_lo_u32_b32{{(_e[0-9]+)?}} v[[mbcnt:[0-9]+]], s[[exec_lo]], 0
106; GCN64: v_mbcnt_hi_u32_b32{{(_e[0-9]+)?}} v[[mbcnt]], s[[exec_hi]], v[[mbcnt]]
107; GCN: v_cmp_eq_u32{{(_e[0-9]+)?}} vcc{{(_lo)?}}, 0, v[[mbcnt]]
108; GCN: s_and_saveexec_b{{32|64}} s[[exec:\[?[0-9:]+\]?]], vcc
109; GCN32: s_bcnt1_i32_b32 s[[popcount:[0-9]+]], s[[exec_lo]]
110; GCN64: s_bcnt1_i32_b64 s[[popcount:[0-9]+]], s{{\[}}[[exec_lo]]:[[exec_hi]]{{\]}}
111; GCN: s_mul_i32 s[[scalar_value:[0-9]+]], s{{[0-9]+}}, s[[popcount]]
112; GCN: v_mov_b32{{(_e[0-9]+)?}} v[[value:[0-9]+]], s[[scalar_value]]
113; GCN: buffer_atomic_sub v[[value]]
114define amdgpu_kernel void @sub_i32_uniform(i32 addrspace(1)* %out, <4 x i32> %inout, i32 %subitive) {
115entry:
116  %old = call i32 @llvm.amdgcn.raw.buffer.atomic.sub(i32 %subitive, <4 x i32> %inout, i32 0, i32 0, i32 0)
117  store i32 %old, i32 addrspace(1)* %out
118  ret void
119}
120
121; GCN-LABEL: sub_i32_varying_vdata:
122; GFX7LESS-NOT: v_mbcnt_lo_u32_b32
123; GFX7LESS-NOT: v_mbcnt_hi_u32_b32
124; GFX7LESS-NOT: s_bcnt1_i32_b64
125; GFX7LESS: buffer_atomic_sub v{{[0-9]+}}
126; DPPCOMB: v_add_u32_dpp
127; DPPCOMB: v_add_u32_dpp
128; GFX8MORE32: v_readlane_b32 s[[scalar_value:[0-9]+]], v{{[0-9]+}}, 31
129; GFX8MORE64: v_readlane_b32 s[[scalar_value:[0-9]+]], v{{[0-9]+}}, 63
130; GFX89: v_mov_b32{{(_e[0-9]+)?}} v[[value:[0-9]+]], s[[scalar_value]]
131; GFX10: s_mov_b32 s[[copy_value:[0-9]+]], s[[scalar_value]]
132; GFX10: v_mov_b32{{(_e[0-9]+)?}} v[[value:[0-9]+]], s[[copy_value]]
133; GFX8MORE: buffer_atomic_sub v[[value]]
134define amdgpu_kernel void @sub_i32_varying_vdata(i32 addrspace(1)* %out, <4 x i32> %inout) {
135entry:
136  %lane = call i32 @llvm.amdgcn.workitem.id.x()
137  %old = call i32 @llvm.amdgcn.raw.buffer.atomic.sub(i32 %lane, <4 x i32> %inout, i32 0, i32 0, i32 0)
138  store i32 %old, i32 addrspace(1)* %out
139  ret void
140}
141
142; GCN-LABEL: sub_i32_varying_offset:
143; GCN-NOT: v_mbcnt_lo_u32_b32
144; GCN-NOT: v_mbcnt_hi_u32_b32
145; GCN-NOT: s_bcnt1_i32_b64
146; GCN: buffer_atomic_sub v{{[0-9]+}}
147define amdgpu_kernel void @sub_i32_varying_offset(i32 addrspace(1)* %out, <4 x i32> %inout) {
148entry:
149  %lane = call i32 @llvm.amdgcn.workitem.id.x()
150  %old = call i32 @llvm.amdgcn.raw.buffer.atomic.sub(i32 1, <4 x i32> %inout, i32 %lane, i32 0, i32 0)
151  store i32 %old, i32 addrspace(1)* %out
152  ret void
153}
154