1; RUN: llc -march=amdgcn -amdgpu-atomic-optimizations=true -verify-machineinstrs < %s | FileCheck -enable-var-scope -check-prefixes=GCN,GCN64,GFX7LESS %s
2; RUN: llc -march=amdgcn -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 -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 -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 -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.struct.buffer.atomic.add(i32, <4 x i32>, i32, i32, i32, i32)
9declare i32 @llvm.amdgcn.struct.buffer.atomic.sub(i32, <4 x i32>, i32, i32, i32, i32)
10
11; Show what the atomic optimization pass will do for struct 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: s_mul_i32 s[[popcount]], s[[popcount]], 5
23; GCN: v_mov_b32{{(_e[0-9]+)?}} v[[value:[0-9]+]], s[[popcount]]
24; GCN: buffer_atomic_add v[[value]]
25define amdgpu_kernel void @add_i32_constant(i32 addrspace(1)* %out, <4 x i32> %inout) {
26entry:
27  %old = call i32 @llvm.amdgcn.struct.buffer.atomic.add(i32 5, <4 x i32> %inout, i32 0, i32 0, i32 0, i32 0)
28  store i32 %old, i32 addrspace(1)* %out
29  ret void
30}
31
32; GCN-LABEL: add_i32_uniform:
33; GCN32: s_mov_b32 s[[exec_lo:[0-9]+]], exec_lo
34; GCN64: s_mov_b64 s{{\[}}[[exec_lo:[0-9]+]]:[[exec_hi:[0-9]+]]{{\]}}, exec
35; GCN: v_mbcnt_lo_u32_b32{{(_e[0-9]+)?}} v[[mbcnt:[0-9]+]], s[[exec_lo]], 0
36; GCN64: v_mbcnt_hi_u32_b32{{(_e[0-9]+)?}} v[[mbcnt]], s[[exec_hi]], v[[mbcnt]]
37; GCN: v_cmp_eq_u32{{(_e[0-9]+)?}} vcc{{(_lo)?}}, 0, v[[mbcnt]]
38; GCN: s_and_saveexec_b{{32|64}} s[[exec:\[?[0-9:]+\]?]], vcc
39; GCN32: s_bcnt1_i32_b32 s[[popcount:[0-9]+]], s[[exec_lo]]
40; GCN64: s_bcnt1_i32_b64 s[[popcount:[0-9]+]], s{{\[}}[[exec_lo]]:[[exec_hi]]{{\]}}
41; GCN: s_mul_i32 s[[scalar_value:[0-9]+]], s{{[0-9]+}}, s[[popcount]]
42; GCN: v_mov_b32{{(_e[0-9]+)?}} v[[value:[0-9]+]], s[[scalar_value]]
43; GCN: buffer_atomic_add v[[value]]
44define amdgpu_kernel void @add_i32_uniform(i32 addrspace(1)* %out, <4 x i32> %inout, i32 %additive) {
45entry:
46  %old = call i32 @llvm.amdgcn.struct.buffer.atomic.add(i32 %additive, <4 x i32> %inout, i32 0, i32 0, i32 0, i32 0)
47  store i32 %old, i32 addrspace(1)* %out
48  ret void
49}
50
51; GCN-LABEL: add_i32_varying_vdata:
52; GFX7LESS-NOT: v_mbcnt_lo_u32_b32
53; GFX7LESS-NOT: v_mbcnt_hi_u32_b32
54; GFX7LESS-NOT: s_bcnt1_i32_b64
55; GFX7LESS: buffer_atomic_add v{{[0-9]+}}
56; DPPCOMB: v_add_u32_dpp
57; DPPCOMB: v_add_u32_dpp
58; GFX8MORE32: v_readlane_b32 s[[scalar_value:[0-9]+]], v{{[0-9]+}}, 31
59; GFX8MORE64: v_readlane_b32 s[[scalar_value:[0-9]+]], v{{[0-9]+}}, 63
60; GFX89: v_mov_b32{{(_e[0-9]+)?}} v[[value:[0-9]+]], s[[scalar_value]]
61; GFX10: s_mov_b32 s[[copy_value:[0-9]+]], s[[scalar_value]]
62; GFX10: v_mov_b32{{(_e[0-9]+)?}} v[[value:[0-9]+]], s[[copy_value]]
63; GFX8MORE: buffer_atomic_add v[[value]]
64define amdgpu_kernel void @add_i32_varying_vdata(i32 addrspace(1)* %out, <4 x i32> %inout) {
65entry:
66  %lane = call i32 @llvm.amdgcn.workitem.id.x()
67  %old = call i32 @llvm.amdgcn.struct.buffer.atomic.add(i32 %lane, <4 x i32> %inout, i32 0, i32 0, i32 0, i32 0)
68  store i32 %old, i32 addrspace(1)* %out
69  ret void
70}
71
72; GCN-LABEL: add_i32_varying_vindex:
73; GCN-NOT: v_mbcnt_lo_u32_b32
74; GCN-NOT: v_mbcnt_hi_u32_b32
75; GCN-NOT: s_bcnt1_i32_b64
76; GCN: buffer_atomic_add v{{[0-9]+}}
77define amdgpu_kernel void @add_i32_varying_vindex(i32 addrspace(1)* %out, <4 x i32> %inout) {
78entry:
79  %lane = call i32 @llvm.amdgcn.workitem.id.x()
80  %old = call i32 @llvm.amdgcn.struct.buffer.atomic.add(i32 1, <4 x i32> %inout, i32 %lane, i32 0, i32 0, i32 0)
81  store i32 %old, i32 addrspace(1)* %out
82  ret void
83}
84
85; GCN-LABEL: add_i32_varying_offset:
86; GCN-NOT: v_mbcnt_lo_u32_b32
87; GCN-NOT: v_mbcnt_hi_u32_b32
88; GCN-NOT: s_bcnt1_i32_b64
89; GCN: buffer_atomic_add v{{[0-9]+}}
90define amdgpu_kernel void @add_i32_varying_offset(i32 addrspace(1)* %out, <4 x i32> %inout) {
91entry:
92  %lane = call i32 @llvm.amdgcn.workitem.id.x()
93  %old = call i32 @llvm.amdgcn.struct.buffer.atomic.add(i32 1, <4 x i32> %inout, i32 0, i32 %lane, i32 0, i32 0)
94  store i32 %old, i32 addrspace(1)* %out
95  ret void
96}
97
98; GCN-LABEL: sub_i32_constant:
99; GCN32: s_mov_b32 s[[exec_lo:[0-9]+]], exec_lo
100; GCN64: s_mov_b64 s{{\[}}[[exec_lo:[0-9]+]]:[[exec_hi:[0-9]+]]{{\]}}, exec
101; GCN: v_mbcnt_lo_u32_b32{{(_e[0-9]+)?}} v[[mbcnt:[0-9]+]], s[[exec_lo]], 0
102; GCN64: v_mbcnt_hi_u32_b32{{(_e[0-9]+)?}} v[[mbcnt]], s[[exec_hi]], v[[mbcnt]]
103; GCN: v_cmp_eq_u32{{(_e[0-9]+)?}} vcc{{(_lo)?}}, 0, v[[mbcnt]]
104; GCN: s_and_saveexec_b{{32|64}} s[[exec:\[?[0-9:]+\]?]], vcc
105; GCN32: s_bcnt1_i32_b32 s[[popcount:[0-9]+]], s[[exec_lo]]
106; GCN64: s_bcnt1_i32_b64 s[[popcount:[0-9]+]], s{{\[}}[[exec_lo]]:[[exec_hi]]{{\]}}
107; GCN: s_mul_i32 s[[popcount]], s[[popcount]], 5
108; GCN: v_mov_b32{{(_e[0-9]+)?}} v[[value:[0-9]+]], s[[popcount]]
109; GCN: buffer_atomic_sub v[[value]]
110define amdgpu_kernel void @sub_i32_constant(i32 addrspace(1)* %out, <4 x i32> %inout) {
111entry:
112  %old = call i32 @llvm.amdgcn.struct.buffer.atomic.sub(i32 5, <4 x i32> %inout, i32 0, i32 0, i32 0, i32 0)
113  store i32 %old, i32 addrspace(1)* %out
114  ret void
115}
116
117; GCN-LABEL: sub_i32_uniform:
118; GCN32: s_mov_b32 s[[exec_lo:[0-9]+]], exec_lo
119; GCN64: s_mov_b64 s{{\[}}[[exec_lo:[0-9]+]]:[[exec_hi:[0-9]+]]{{\]}}, exec
120; GCN: v_mbcnt_lo_u32_b32{{(_e[0-9]+)?}} v[[mbcnt:[0-9]+]], s[[exec_lo]], 0
121; GCN64: v_mbcnt_hi_u32_b32{{(_e[0-9]+)?}} v[[mbcnt]], s[[exec_hi]], v[[mbcnt]]
122; GCN: v_cmp_eq_u32{{(_e[0-9]+)?}} vcc{{(_lo)?}}, 0, v[[mbcnt]]
123; GCN: s_and_saveexec_b{{32|64}} s[[exec:\[?[0-9:]+\]?]], vcc
124; GCN32: s_bcnt1_i32_b32 s[[popcount:[0-9]+]], s[[exec_lo]]
125; GCN64: s_bcnt1_i32_b64 s[[popcount:[0-9]+]], s{{\[}}[[exec_lo]]:[[exec_hi]]{{\]}}
126; GCN: s_mul_i32 s[[scalar_value:[0-9]+]], s{{[0-9]+}}, s[[popcount]]
127; GCN: v_mov_b32{{(_e[0-9]+)?}} v[[value:[0-9]+]], s[[scalar_value]]
128; GCN: buffer_atomic_sub v[[value]]
129define amdgpu_kernel void @sub_i32_uniform(i32 addrspace(1)* %out, <4 x i32> %inout, i32 %subitive) {
130entry:
131  %old = call i32 @llvm.amdgcn.struct.buffer.atomic.sub(i32 %subitive, <4 x i32> %inout, i32 0, i32 0, i32 0, i32 0)
132  store i32 %old, i32 addrspace(1)* %out
133  ret void
134}
135
136; GCN-LABEL: sub_i32_varying_vdata:
137; GFX7LESS-NOT: v_mbcnt_lo_u32_b32
138; GFX7LESS-NOT: v_mbcnt_hi_u32_b32
139; GFX7LESS-NOT: s_bcnt1_i32_b64
140; GFX7LESS: buffer_atomic_sub v{{[0-9]+}}
141; DPPCOMB: v_add_u32_dpp
142; DPPCOMB: v_add_u32_dpp
143; GFX8MORE32: v_readlane_b32 s[[scalar_value:[0-9]+]], v{{[0-9]+}}, 31
144; GFX8MORE64: v_readlane_b32 s[[scalar_value:[0-9]+]], v{{[0-9]+}}, 63
145; GFX89: v_mov_b32{{(_e[0-9]+)?}} v[[value:[0-9]+]], s[[scalar_value]]
146; GFX10: s_mov_b32 s[[copy_value:[0-9]+]], s[[scalar_value]]
147; GFX10: v_mov_b32{{(_e[0-9]+)?}} v[[value:[0-9]+]], s[[copy_value]]
148; GFX8MORE: buffer_atomic_sub v[[value]]
149define amdgpu_kernel void @sub_i32_varying_vdata(i32 addrspace(1)* %out, <4 x i32> %inout) {
150entry:
151  %lane = call i32 @llvm.amdgcn.workitem.id.x()
152  %old = call i32 @llvm.amdgcn.struct.buffer.atomic.sub(i32 %lane, <4 x i32> %inout, i32 0, i32 0, i32 0, i32 0)
153  store i32 %old, i32 addrspace(1)* %out
154  ret void
155}
156
157; GCN-LABEL: sub_i32_varying_vindex:
158; GCN-NOT: v_mbcnt_lo_u32_b32
159; GCN-NOT: v_mbcnt_hi_u32_b32
160; GCN-NOT: s_bcnt1_i32_b64
161; GCN: buffer_atomic_sub v{{[0-9]+}}
162define amdgpu_kernel void @sub_i32_varying_vindex(i32 addrspace(1)* %out, <4 x i32> %inout) {
163entry:
164  %lane = call i32 @llvm.amdgcn.workitem.id.x()
165  %old = call i32 @llvm.amdgcn.struct.buffer.atomic.sub(i32 1, <4 x i32> %inout, i32 %lane, i32 0, i32 0, i32 0)
166  store i32 %old, i32 addrspace(1)* %out
167  ret void
168}
169
170; GCN-LABEL: sub_i32_varying_offset:
171; GCN-NOT: v_mbcnt_lo_u32_b32
172; GCN-NOT: v_mbcnt_hi_u32_b32
173; GCN-NOT: s_bcnt1_i32_b64
174; GCN: buffer_atomic_sub v{{[0-9]+}}
175define amdgpu_kernel void @sub_i32_varying_offset(i32 addrspace(1)* %out, <4 x i32> %inout) {
176entry:
177  %lane = call i32 @llvm.amdgcn.workitem.id.x()
178  %old = call i32 @llvm.amdgcn.struct.buffer.atomic.sub(i32 1, <4 x i32> %inout, i32 0, i32 %lane, i32 0, i32 0)
179  store i32 %old, i32 addrspace(1)* %out
180  ret void
181}
182