1; RUN: llc -amdgpu-scalarize-global-loads=false -march=amdgcn -mcpu=tahiti -verify-machineinstrs < %s | FileCheck -enable-var-scope -check-prefixes=GCN,SI,SIVI %s
2; RUN: llc -amdgpu-scalarize-global-loads=false -march=amdgcn -mcpu=fiji -mattr=-flat-for-global -verify-machineinstrs < %s | FileCheck -enable-var-scope -check-prefixes=GCN,VI,GFX89,SIVI %s
3; RUN: llc -amdgpu-scalarize-global-loads=false -march=amdgcn -mcpu=gfx900 -mattr=-flat-for-global -verify-machineinstrs < %s | FileCheck -enable-var-scope -check-prefixes=GCN,GFX89,GFX9 %s
4
5declare half @llvm.minnum.f16(half %a, half %b)
6declare <2 x half> @llvm.minnum.v2f16(<2 x half> %a, <2 x half> %b)
7declare <3 x half> @llvm.minnum.v3f16(<3 x half> %a, <3 x half> %b)
8declare <4 x half> @llvm.minnum.v4f16(<4 x half> %a, <4 x half> %b)
9
10; GCN-LABEL: {{^}}minnum_f16:
11; GCN: buffer_load_ushort v[[A_F16:[0-9]+]]
12; GCN: buffer_load_ushort v[[B_F16:[0-9]+]]
13; SI:  v_cvt_f32_f16_e32 v[[A_F32:[0-9]+]], v[[A_F16]]
14; SI:  v_cvt_f32_f16_e32 v[[B_F32:[0-9]+]], v[[B_F16]]
15; SI:  v_min_f32_e32 v[[R_F32:[0-9]+]], v[[A_F32]], v[[B_F32]]
16; SI:  v_cvt_f16_f32_e32 v[[R_F16:[0-9]+]], v[[R_F32]]
17; GFX89:  v_min_f16_e32 v[[R_F16:[0-9]+]], v[[A_F16]], v[[B_F16]]
18; GCN: buffer_store_short v[[R_F16]]
19; GCN: s_endpgm
20define amdgpu_kernel void @minnum_f16(
21    half addrspace(1)* %r,
22    half addrspace(1)* %a,
23    half addrspace(1)* %b) {
24entry:
25  %a.val = load volatile half, half addrspace(1)* %a
26  %b.val = load volatile half, half addrspace(1)* %b
27  %r.val = call half @llvm.minnum.f16(half %a.val, half %b.val)
28  store half %r.val, half addrspace(1)* %r
29  ret void
30}
31
32; GCN-LABEL: {{^}}minnum_f16_imm_a:
33; GCN: buffer_load_ushort v[[B_F16:[0-9]+]]
34; SI:  v_cvt_f32_f16_e32 v[[B_F32:[0-9]+]], v[[B_F16]]
35; SI:  v_min_f32_e32 v[[R_F32:[0-9]+]], 0x40400000, v[[B_F32]]
36; SI:  v_cvt_f16_f32_e32 v[[R_F16:[0-9]+]], v[[R_F32]]
37; GFX89:  v_min_f16_e32 v[[R_F16:[0-9]+]], 0x4200, v[[B_F16]]
38; GCN: buffer_store_short v[[R_F16]]
39; GCN: s_endpgm
40define amdgpu_kernel void @minnum_f16_imm_a(
41    half addrspace(1)* %r,
42    half addrspace(1)* %b) {
43entry:
44  %b.val = load half, half addrspace(1)* %b
45  %r.val = call half @llvm.minnum.f16(half 3.0, half %b.val)
46  store half %r.val, half addrspace(1)* %r
47  ret void
48}
49
50; GCN-LABEL: {{^}}minnum_f16_imm_b:
51; GCN: buffer_load_ushort v[[A_F16:[0-9]+]]
52; SI:  v_cvt_f32_f16_e32 v[[A_F32:[0-9]+]], v[[A_F16]]
53; SI:  v_min_f32_e32 v[[R_F32:[0-9]+]], 4.0, v[[A_F32]]
54; SI:  v_cvt_f16_f32_e32 v[[R_F16:[0-9]+]], v[[R_F32]]
55; GFX89:  v_min_f16_e32 v[[R_F16:[0-9]+]], 4.0, v[[A_F16]]
56; GCN: buffer_store_short v[[R_F16]]
57; GCN: s_endpgm
58define amdgpu_kernel void @minnum_f16_imm_b(
59    half addrspace(1)* %r,
60    half addrspace(1)* %a) {
61entry:
62  %a.val = load half, half addrspace(1)* %a
63  %r.val = call half @llvm.minnum.f16(half %a.val, half 4.0)
64  store half %r.val, half addrspace(1)* %r
65  ret void
66}
67
68; GCN-LABEL: {{^}}minnum_v2f16:
69; GCN: buffer_load_dword v[[B_V2_F16:[0-9]+]]
70; GCN: buffer_load_dword v[[A_V2_F16:[0-9]+]]
71
72; SI: v_lshrrev_b32_e32 v[[B_F16_1:[0-9]+]], 16, v[[B_V2_F16]]
73; SI: v_lshrrev_b32_e32 v[[A_F16_1:[0-9]+]], 16, v[[A_V2_F16]]
74; SI: v_cvt_f32_f16_e32 v[[B_F32_1:[0-9]+]], v[[B_F16_1]]
75; SI: v_cvt_f32_f16_e32 v[[A_F32_1:[0-9]+]], v[[A_F16_1]]
76; SI: v_cvt_f32_f16_e32 v[[B_F32_0:[0-9]+]], v[[B_V2_F16]]
77; SI: v_cvt_f32_f16_e32 v[[A_F32_0:[0-9]+]], v[[A_V2_F16]]
78; SI-DAG: v_min_f32_e32 v[[R_F32_0:[0-9]+]], v[[A_F32_0]], v[[B_F32_0]]
79; SI-DAG: v_min_f32_e32 v[[R_F32_1:[0-9]+]], v[[A_F32_1]], v[[B_F32_1]]
80; SI-DAG: v_cvt_f16_f32_e32 v[[R_F16_0:[0-9]+]], v[[R_F32_0]]
81; SI-DAG: v_cvt_f16_f32_e32 v[[R_F16_1:[0-9]+]], v[[R_F32_1]]
82; SI:     v_lshlrev_b32_e32 v[[R_F16_HI:[0-9]+]], 16, v[[R_F16_1]]
83; SI-NOT: and
84; SI: v_or_b32_e32 v[[R_V2_F16:[0-9]+]], v[[R_F16_0]], v[[R_F16_HI]]
85
86; VI-DAG: v_min_f16_e32 v[[R_F16_0:[0-9]+]], v[[A_V2_F16]], v[[B_V2_F16]]
87; VI-DAG: v_min_f16_sdwa v[[R_F16_1:[0-9]+]], v[[A_V2_F16]], v[[B_V2_F16]] dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:WORD_1
88; VI-NOT: and
89; VI:    v_or_b32_e32 v[[R_V2_F16:[0-9]+]], v[[R_F16_0]], v[[R_F16_1]]
90
91; GFX9: v_pk_min_f16 v[[R_V2_F16:[0-9]+]], v[[B_V2_F16]], v[[A_V2_F16]]
92
93; GCN: buffer_store_dword v[[R_V2_F16]]
94define amdgpu_kernel void @minnum_v2f16(
95    <2 x half> addrspace(1)* %r,
96    <2 x half> addrspace(1)* %a,
97    <2 x half> addrspace(1)* %b) {
98entry:
99  %a.val = load <2 x half>, <2 x half> addrspace(1)* %a
100  %b.val = load <2 x half>, <2 x half> addrspace(1)* %b
101  %r.val = call <2 x half> @llvm.minnum.v2f16(<2 x half> %a.val, <2 x half> %b.val)
102  store <2 x half> %r.val, <2 x half> addrspace(1)* %r
103  ret void
104}
105
106; GCN-LABEL: {{^}}minnum_v2f16_imm_a:
107; GCN-DAG: buffer_load_dword v[[B_V2_F16:[0-9]+]]
108; SI-DAG: v_cvt_f32_f16_e32 v[[B_F32_0:[0-9]+]], v[[B_V2_F16]]
109; SI-DAG: v_lshrrev_b32_e32 v[[B_F16_1:[0-9]+]], 16, v[[B_V2_F16]]
110; SI-DAG: v_cvt_f32_f16_e32 v[[B_F32_1:[0-9]+]], v[[B_F16_1]]
111; SI-DAG: v_min_f32_e32 v[[R_F32_0:[0-9]+]], 0x40400000, v[[B_F32_0]]
112; SI-DAG: v_cvt_f16_f32_e32 v[[R_F16_0:[0-9]+]], v[[R_F32_0]]
113; SI-DAG: v_min_f32_e32 v[[R_F32_1:[0-9]+]], 4.0, v[[B_F32_1]]
114; SI-DAG:  v_cvt_f16_f32_e32 v[[R_F16_1:[0-9]+]], v[[R_F32_1]]
115; VI-DAG:  v_mov_b32_e32 [[CONST4:v[0-9]+]], 0x4400
116; VI-DAG:  v_min_f16_sdwa v[[R_F16_HI:[0-9]+]], v[[B_V2_F16]], [[CONST4]] dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
117; VI-DAG:  v_min_f16_e32 v[[R_F16_0:[0-9]+]], 0x4200, v[[B_V2_F16]]
118
119; SI-DAG: v_lshlrev_b32_e32 v[[R_F16_HI:[0-9]+]], 16, v[[R_F16_1]]
120; SIVI-NOT: and
121; SIVI:  v_or_b32_e32 v[[R_V2_F16:[0-9]+]], v[[R_F16_0]], v[[R_F16_HI]]
122
123
124; GFX9: s_mov_b32 [[K:s[0-9]+]], 0x44004200
125; GFX9: v_pk_min_f16 v[[R_V2_F16:[0-9]+]], v[[B_V2_F16]], [[K]]
126
127; GCN: buffer_store_dword v[[R_V2_F16]]
128define amdgpu_kernel void @minnum_v2f16_imm_a(
129    <2 x half> addrspace(1)* %r,
130    <2 x half> addrspace(1)* %b) {
131entry:
132  %b.val = load <2 x half>, <2 x half> addrspace(1)* %b
133  %r.val = call <2 x half> @llvm.minnum.v2f16(<2 x half> <half 3.0, half 4.0>, <2 x half> %b.val)
134  store <2 x half> %r.val, <2 x half> addrspace(1)* %r
135  ret void
136}
137
138; GCN-LABEL: {{^}}minnum_v2f16_imm_b:
139; GCN-DAG: buffer_load_dword v[[A_V2_F16:[0-9]+]]
140; SI-DAG: v_cvt_f32_f16_e32 v[[A_F32_0:[0-9]+]], v[[A_V2_F16]]
141; SI-DAG: v_lshrrev_b32_e32 v[[A_F16_1:[0-9]+]], 16, v[[A_V2_F16]]
142; SI-DAG: v_cvt_f32_f16_e32 v[[A_F32_1:[0-9]+]], v[[A_F16_1]]
143; SI-DAG: v_min_f32_e32 v[[R_F32_0:[0-9]+]], 4.0, v[[A_F32_0]]
144; SI-DAG: v_cvt_f16_f32_e32 v[[R_F16_0:[0-9]+]], v[[R_F32_0]]
145; SI-DAG: v_min_f32_e32 v[[R_F32_1:[0-9]+]], 0x40400000, v[[A_F32_1]]
146; SI-DAG: v_cvt_f16_f32_e32 v[[R_F16_1:[0-9]+]], v[[R_F32_1]]
147
148; VI-DAG:  v_mov_b32_e32 [[CONST3:v[0-9]+]], 0x4200
149; VI-DAG:  v_min_f16_sdwa v[[R_F16_HI:[0-9]+]], v[[A_V2_F16]], [[CONST3]] dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
150; VI-DAG:  v_min_f16_e32 v[[R_F16_0:[0-9]+]], 4.0, v[[A_V2_F16]]
151
152; SI-DAG: v_lshlrev_b32_e32 v[[R_F16_HI:[0-9]+]], 16, v[[R_F16_1]]
153
154
155; SIVI-NOT: and
156; SIVI: v_or_b32_e32 v[[R_V2_F16:[0-9]+]], v[[R_F16_0]], v[[R_F16_HI]]
157
158; GFX9: s_mov_b32 [[K:s[0-9]+]], 0x42004400
159; GFX9: v_pk_min_f16 v[[R_V2_F16:[0-9]+]], v[[A_V2_F16]], [[K]]
160
161; GCN: buffer_store_dword v[[R_V2_F16]]
162define amdgpu_kernel void @minnum_v2f16_imm_b(
163    <2 x half> addrspace(1)* %r,
164    <2 x half> addrspace(1)* %a) {
165entry:
166  %a.val = load <2 x half>, <2 x half> addrspace(1)* %a
167  %r.val = call <2 x half> @llvm.minnum.v2f16(<2 x half> %a.val, <2 x half> <half 4.0, half 3.0>)
168  store <2 x half> %r.val, <2 x half> addrspace(1)* %r
169  ret void
170}
171
172; FIXME: Scalarize with undef half
173; GCN-LABEL: {{^}}minnum_v3f16:
174; GFX9: v_pk_min_f16
175; GFX9: v_pk_min_f16
176define amdgpu_kernel void @minnum_v3f16(
177    <3 x half> addrspace(1)* %r,
178    <3 x half> addrspace(1)* %a,
179    <3 x half> addrspace(1)* %b) {
180entry:
181  %a.val = load <3 x half>, <3 x half> addrspace(1)* %a
182  %b.val = load <3 x half>, <3 x half> addrspace(1)* %b
183  %r.val = call <3 x half> @llvm.minnum.v3f16(<3 x half> %a.val, <3 x half> %b.val)
184  store <3 x half> %r.val, <3 x half> addrspace(1)* %r
185  ret void
186}
187
188; GCN-LABEL: {{^}}minnum_v4f16:
189; GFX89: buffer_load_dwordx2 v{{\[}}[[A_LO:[0-9]+]]:[[A_HI:[0-9]+]]{{\]}}
190; GFX89: buffer_load_dwordx2 v{{\[}}[[B_LO:[0-9]+]]:[[B_HI:[0-9]+]]{{\]}}
191; GFX9-DAG: v_pk_min_f16 v[[MIN_LO:[0-9]+]], v[[B_LO]], v[[A_LO]]
192; GFX9-DAG: v_pk_min_f16 v[[MIN_HI:[0-9]+]], v[[B_HI]], v[[A_HI]]
193; GFX9: buffer_store_dwordx2 v{{\[}}[[MIN_LO]]:[[MIN_HI]]{{\]}}
194define amdgpu_kernel void @minnum_v4f16(
195    <4 x half> addrspace(1)* %r,
196    <4 x half> addrspace(1)* %a,
197    <4 x half> addrspace(1)* %b) {
198entry:
199  %a.val = load <4 x half>, <4 x half> addrspace(1)* %a
200  %b.val = load <4 x half>, <4 x half> addrspace(1)* %b
201  %r.val = call <4 x half> @llvm.minnum.v4f16(<4 x half> %a.val, <4 x half> %b.val)
202  store <4 x half> %r.val, <4 x half> addrspace(1)* %r
203  ret void
204}
205
206; GCN-LABEL: {{^}}fmin_v4f16_imm_a:
207; GFX89-DAG: buffer_load_dwordx2 v{{\[}}[[A_LO:[0-9]+]]:[[A_HI:[0-9]+]]{{\]}}
208; GFX9-DAG: s_mov_b32 [[K1:s[0-9]+]], 0x44004200
209; GFX9-DAG: s_mov_b32 [[K0:s[0-9]+]], 0x40004800
210
211; GFX9-DAG: v_pk_min_f16 v[[MIN_LO:[0-9]+]], v[[A_LO]], [[K0]]
212; GFX9-DAG: v_pk_min_f16 v[[MIN_HI:[0-9]+]], v[[A_HI]], [[K1]]
213; GFX9: buffer_store_dwordx2 v{{\[}}[[MIN_LO]]:[[MIN_HI]]{{\]}}
214
215; VI-DAG: v_mov_b32_e32 [[K2:v[0-9]+]], 0x4000
216; VI-DAG: v_mov_b32_e32 [[K4:v[0-9]+]], 0x4400
217
218; VI-DAG: v_min_f16_sdwa v[[MIN_HI_HI:[0-9]+]], v[[A_HI]], [[K4]] dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
219; VI-DAG: v_min_f16_e32 v[[MIN_HI_LO:[0-9]+]], 0x4200, v[[A_HI]]
220; VI-DAG: v_min_f16_sdwa v[[MIN_LO_HI:[0-9]+]], v[[A_LO]], [[K2]] dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
221; VI-DAG: v_min_f16_e32 v[[MIN_LO_LO:[0-9]+]], 0x4800, v[[A_LO]]
222
223; VI-DAG: v_or_b32_e32 v[[OR0:[0-9]+]], v[[MIN_LO_LO]], v[[MIN_LO_HI]]
224; VI-DAG: v_or_b32_e32 v[[OR1:[0-9]+]], v[[MIN_HI_LO]], v[[MIN_HI_HI]]
225
226; VI: buffer_store_dwordx2 v{{\[}}[[OR0]]:[[OR1]]{{\]}}
227define amdgpu_kernel void @fmin_v4f16_imm_a(
228    <4 x half> addrspace(1)* %r,
229    <4 x half> addrspace(1)* %b) {
230entry:
231  %b.val = load <4 x half>, <4 x half> addrspace(1)* %b
232  %r.val = call <4 x half> @llvm.minnum.v4f16(<4 x half> <half 8.0, half 2.0, half 3.0, half 4.0>, <4 x half> %b.val)
233  store <4 x half> %r.val, <4 x half> addrspace(1)* %r
234  ret void
235}
236