1; RUN:  llc -amdgpu-scalarize-global-loads=false  -march=amdgcn -mcpu=tahiti -mattr=-fp64-fp16-denormals -verify-machineinstrs < %s | FileCheck -allow-deprecated-dag-overlap -check-prefix=GCN -check-prefix=SI %s
2; RUN:  llc -amdgpu-scalarize-global-loads=false  -march=amdgcn -mcpu=fiji -mattr=-fp64-fp16-denormals,-flat-for-global -verify-machineinstrs < %s | FileCheck -allow-deprecated-dag-overlap -check-prefix=GCN -check-prefix=VI %s
3
4; GCN-LABEL: {{^}}mac_f16:
5; GCN: {{buffer|flat}}_load_ushort v[[A_F16:[0-9]+]]
6; GCN: {{buffer|flat}}_load_ushort v[[B_F16:[0-9]+]]
7; GCN: {{buffer|flat}}_load_ushort v[[C_F16:[0-9]+]]
8; SI:  v_cvt_f32_f16_e32 v[[A_F32:[0-9]+]], v[[A_F16]]
9; SI:  v_cvt_f32_f16_e32 v[[B_F32:[0-9]+]], v[[B_F16]]
10; SI:  v_cvt_f32_f16_e32 v[[C_F32:[0-9]+]], v[[C_F16]]
11; SI:  v_mac_f32_e32 v[[C_F32]], v[[A_F32]], v[[B_F32]]
12; SI:  v_cvt_f16_f32_e32 v[[R_F16:[0-9]+]], v[[C_F32]]
13; SI:  buffer_store_short v[[R_F16]]
14; VI:  v_mac_f16_e32 v[[C_F16]], v[[A_F16]], v[[B_F16]]
15; VI:  buffer_store_short v[[C_F16]]
16; GCN: s_endpgm
17define amdgpu_kernel void @mac_f16(
18    half addrspace(1)* %r,
19    half addrspace(1)* %a,
20    half addrspace(1)* %b,
21    half addrspace(1)* %c) #0 {
22entry:
23  %a.val = load half, half addrspace(1)* %a
24  %b.val = load half, half addrspace(1)* %b
25  %c.val = load half, half addrspace(1)* %c
26
27  %t.val = fmul half %a.val, %b.val
28  %r.val = fadd half %t.val, %c.val
29
30  store half %r.val, half addrspace(1)* %r
31  ret void
32}
33
34; GCN-LABEL: {{^}}mac_f16_same_add:
35; SI:  v_mad_f32 v{{[0-9]}}, v{{[0-9]+}}, v{{[0-9]+}}, [[ADD:v[0-9]+]]
36; SI:  v_mac_f32_e32 [[ADD]], v{{[0-9]+}}, v{{[0-9]+}}
37
38; VI:  v_mad_f16 v{{[0-9]}}, v{{[0-9]+}}, v{{[0-9]+}}, [[ADD:v[0-9]+]]
39; VI:  v_mac_f16_e32 [[ADD]], v{{[0-9]+}}, v{{[0-9]+}}
40; GCN: s_endpgm
41define amdgpu_kernel void @mac_f16_same_add(
42    half addrspace(1)* %r0,
43    half addrspace(1)* %r1,
44    half addrspace(1)* %a,
45    half addrspace(1)* %b,
46    half addrspace(1)* %c,
47    half addrspace(1)* %d,
48    half addrspace(1)* %e) #0 {
49entry:
50  %a.val = load half, half addrspace(1)* %a
51  %b.val = load half, half addrspace(1)* %b
52  %c.val = load half, half addrspace(1)* %c
53  %d.val = load half, half addrspace(1)* %d
54  %e.val = load half, half addrspace(1)* %e
55
56  %t0.val = fmul half %a.val, %b.val
57  %r0.val = fadd half %t0.val, %c.val
58
59  %t1.val = fmul half %d.val, %e.val
60  %r1.val = fadd half %t1.val, %c.val
61
62  store half %r0.val, half addrspace(1)* %r0
63  store half %r1.val, half addrspace(1)* %r1
64  ret void
65}
66
67; GCN-LABEL: {{^}}mac_f16_neg_a:
68; SI: v_cvt_f32_f16_e32 [[CVT_A:v[0-9]+]], v{{[0-9]+}}
69; SI: v_cvt_f32_f16_e32 [[CVT_B:v[0-9]+]], v{{[0-9]+}}
70; SI: v_cvt_f32_f16_e32 [[CVT_C:v[0-9]+]], v{{[0-9]+}}
71; SI: v_mad_f32 v{{[0-9]+}}, -[[CVT_A]], [[CVT_B]], [[CVT_C]]
72
73; VI-NOT: v_mac_f16
74; VI:     v_mad_f16 v{{[0-9]+}}, -v{{[0-9]+}}, v{{[0-9]+}}, v{{[0-9]+}}
75; GCN:    s_endpgm
76define amdgpu_kernel void @mac_f16_neg_a(
77    half addrspace(1)* %r,
78    half addrspace(1)* %a,
79    half addrspace(1)* %b,
80    half addrspace(1)* %c) #0 {
81entry:
82  %a.val = load half, half addrspace(1)* %a
83  %b.val = load half, half addrspace(1)* %b
84  %c.val = load half, half addrspace(1)* %c
85
86  %a.neg = fsub half -0.0, %a.val
87  %t.val = fmul half %a.neg, %b.val
88  %r.val = fadd half %t.val, %c.val
89
90  store half %r.val, half addrspace(1)* %r
91  ret void
92}
93
94; GCN-LABEL: {{^}}mac_f16_neg_b:
95; SI: v_cvt_f32_f16_e32 [[CVT_A:v[0-9]+]], v{{[0-9]+}}
96; SI: v_cvt_f32_f16_e32 [[CVT_B:v[0-9]+]], v{{[0-9]+}}
97; SI: v_cvt_f32_f16_e32 [[CVT_C:v[0-9]+]], v{{[0-9]+}}
98; SI: v_mad_f32 v{{[0-9]+}}, -[[CVT_A]], [[CVT_B]], [[CVT_C]]
99
100; VI-NOT: v_mac_f16
101; VI:     v_mad_f16 v{{[0-9]+}}, -v{{[0-9]+}}, v{{[0-9]+}}, v{{[0-9]+}}
102; GCN:    s_endpgm
103define amdgpu_kernel void @mac_f16_neg_b(
104    half addrspace(1)* %r,
105    half addrspace(1)* %a,
106    half addrspace(1)* %b,
107    half addrspace(1)* %c) #0 {
108entry:
109  %a.val = load half, half addrspace(1)* %a
110  %b.val = load half, half addrspace(1)* %b
111  %c.val = load half, half addrspace(1)* %c
112
113  %b.neg = fsub half -0.0, %b.val
114  %t.val = fmul half %a.val, %b.neg
115  %r.val = fadd half %t.val, %c.val
116
117  store half %r.val, half addrspace(1)* %r
118  ret void
119}
120
121; GCN-LABEL: {{^}}mac_f16_neg_c:
122; SI: v_cvt_f32_f16_e32
123; SI: v_cvt_f32_f16_e32
124; SI: v_cvt_f32_f16_e32
125; SI: v_mad_f32 v{{[0-9]+}}, v{{[0-9]+}}, v{{[0-9]+}}, -v{{[0-9]+}}
126
127; VI-NOT: v_mac_f16
128; VI:     v_mad_f16 v{{[0-9]+}}, v{{[0-9]+}}, v{{[0-9]+}}, -v{{[0-9]+}}
129; GCN:    s_endpgm
130define amdgpu_kernel void @mac_f16_neg_c(
131    half addrspace(1)* %r,
132    half addrspace(1)* %a,
133    half addrspace(1)* %b,
134    half addrspace(1)* %c) #0 {
135entry:
136  %a.val = load half, half addrspace(1)* %a
137  %b.val = load half, half addrspace(1)* %b
138  %c.val = load half, half addrspace(1)* %c
139
140  %c.neg = fsub half -0.0, %c.val
141  %t.val = fmul half %a.val, %b.val
142  %r.val = fadd half %t.val, %c.neg
143
144  store half %r.val, half addrspace(1)* %r
145  ret void
146}
147
148; GCN-LABEL: {{^}}mac_f16_neg_a_safe_fp_math:
149; SI:  v_sub_f32_e32 v[[NEG_A:[0-9]+]], 0, v{{[0-9]+}}
150; SI:  v_mac_f32_e32 v{{[0-9]+}}, v[[NEG_A]], v{{[0-9]+}}
151; VI:  v_sub_f16_e32 v[[NEG_A:[0-9]+]], 0, v{{[0-9]+}}
152; VI:  v_mac_f16_e32 v{{[0-9]+}}, v[[NEG_A]], v{{[0-9]+}}
153; GCN: s_endpgm
154define amdgpu_kernel void @mac_f16_neg_a_safe_fp_math(
155    half addrspace(1)* %r,
156    half addrspace(1)* %a,
157    half addrspace(1)* %b,
158    half addrspace(1)* %c) #0 {
159entry:
160  %a.val = load half, half addrspace(1)* %a
161  %b.val = load half, half addrspace(1)* %b
162  %c.val = load half, half addrspace(1)* %c
163
164  %a.neg = fsub half 0.0, %a.val
165  %t.val = fmul half %a.neg, %b.val
166  %r.val = fadd half %t.val, %c.val
167
168  store half %r.val, half addrspace(1)* %r
169  ret void
170}
171
172; GCN-LABEL: {{^}}mac_f16_neg_b_safe_fp_math:
173; SI:  v_sub_f32_e32 v[[NEG_A:[0-9]+]], 0, v{{[0-9]+}}
174; SI:  v_mac_f32_e32 v{{[0-9]+}}, v{{[0-9]+}}, v[[NEG_A]]
175; VI:  v_sub_f16_e32 v[[NEG_A:[0-9]+]], 0, v{{[0-9]+}}
176; VI:  v_mac_f16_e32 v{{[0-9]+}}, v{{[0-9]+}}, v[[NEG_A]]
177; GCN: s_endpgm
178define amdgpu_kernel void @mac_f16_neg_b_safe_fp_math(
179    half addrspace(1)* %r,
180    half addrspace(1)* %a,
181    half addrspace(1)* %b,
182    half addrspace(1)* %c) #0 {
183entry:
184  %a.val = load half, half addrspace(1)* %a
185  %b.val = load half, half addrspace(1)* %b
186  %c.val = load half, half addrspace(1)* %c
187
188  %b.neg = fsub half 0.0, %b.val
189  %t.val = fmul half %a.val, %b.neg
190  %r.val = fadd half %t.val, %c.val
191
192  store half %r.val, half addrspace(1)* %r
193  ret void
194}
195
196; GCN-LABEL: {{^}}mac_f16_neg_c_safe_fp_math:
197; SI:  v_sub_f32_e32 v[[NEG_A:[0-9]+]], 0, v{{[0-9]+}}
198; SI:  v_mac_f32_e32 v[[NEG_A]], v{{[0-9]+}}, v{{[0-9]+}}
199; VI:  v_sub_f16_e32 v[[NEG_A:[0-9]+]], 0, v{{[0-9]+}}
200; VI:  v_mac_f16_e32 v[[NEG_A]], v{{[0-9]+}}, v{{[0-9]+}}
201; GCN: s_endpgm
202define amdgpu_kernel void @mac_f16_neg_c_safe_fp_math(
203    half addrspace(1)* %r,
204    half addrspace(1)* %a,
205    half addrspace(1)* %b,
206    half addrspace(1)* %c) #0 {
207entry:
208  %a.val = load half, half addrspace(1)* %a
209  %b.val = load half, half addrspace(1)* %b
210  %c.val = load half, half addrspace(1)* %c
211
212  %c.neg = fsub half 0.0, %c.val
213  %t.val = fmul half %a.val, %b.val
214  %r.val = fadd half %t.val, %c.neg
215
216  store half %r.val, half addrspace(1)* %r
217  ret void
218}
219
220; GCN-LABEL: {{^}}mac_f16_neg_a_nsz_fp_math:
221; SI: v_cvt_f32_f16_e32 [[CVT_A:v[0-9]+]], v{{[0-9]+}}
222; SI: v_cvt_f32_f16_e32 [[CVT_B:v[0-9]+]], v{{[0-9]+}}
223; SI: v_cvt_f32_f16_e32 [[CVT_C:v[0-9]+]], v{{[0-9]+}}
224; SI: v_mad_f32 v{{[0-9]+}}, -[[CVT_A]], [[CVT_B]], [[CVT_C]]
225
226; VI-NOT: v_mac_f16
227; VI:     v_mad_f16 v{{[0-9]+}}, -v{{[0-9]+}}, v{{[0-9]+}}, v{{[0-9]}}
228; GCN:    s_endpgm
229define amdgpu_kernel void @mac_f16_neg_a_nsz_fp_math(
230    half addrspace(1)* %r,
231    half addrspace(1)* %a,
232    half addrspace(1)* %b,
233    half addrspace(1)* %c) #1 {
234entry:
235  %a.val = load half, half addrspace(1)* %a
236  %b.val = load half, half addrspace(1)* %b
237  %c.val = load half, half addrspace(1)* %c
238
239  %a.neg = fsub half 0.0, %a.val
240  %t.val = fmul half %a.neg, %b.val
241  %r.val = fadd half %t.val, %c.val
242
243  store half %r.val, half addrspace(1)* %r
244  ret void
245}
246
247; GCN-LABEL: {{^}}mac_f16_neg_b_nsz_fp_math:
248; SI: v_cvt_f32_f16_e32 [[CVT_A:v[0-9]+]], v{{[0-9]+}}
249; SI: v_cvt_f32_f16_e32 [[CVT_B:v[0-9]+]], v{{[0-9]+}}
250; SI: v_cvt_f32_f16_e32 [[CVT_C:v[0-9]+]], v{{[0-9]+}}
251; SI: v_mad_f32 v{{[0-9]+}}, -[[CVT_A]], [[CVT_B]], [[CVT_C]]
252
253; VI-NOT: v_mac_f16
254; VI:     v_mad_f16 v{{[0-9]+}}, -v{{[0-9]+}}, v{{[0-9]+}}, v{{[0-9]}}
255; GCN:    s_endpgm
256define amdgpu_kernel void @mac_f16_neg_b_nsz_fp_math(
257    half addrspace(1)* %r,
258    half addrspace(1)* %a,
259    half addrspace(1)* %b,
260    half addrspace(1)* %c) #1 {
261entry:
262  %a.val = load half, half addrspace(1)* %a
263  %b.val = load half, half addrspace(1)* %b
264  %c.val = load half, half addrspace(1)* %c
265
266  %b.neg = fsub half 0.0, %b.val
267  %t.val = fmul half %a.val, %b.neg
268  %r.val = fadd half %t.val, %c.val
269
270  store half %r.val, half addrspace(1)* %r
271  ret void
272}
273
274; GCN-LABEL: {{^}}mac_f16_neg_c_nsz_fp_math:
275; SI: v_cvt_f32_f16_e32 [[CVT_A:v[0-9]+]], v{{[0-9]+}}
276; SI: v_cvt_f32_f16_e32 [[CVT_B:v[0-9]+]], v{{[0-9]+}}
277; SI: v_cvt_f32_f16_e32 [[CVT_C:v[0-9]+]], v{{[0-9]+}}
278; SI: v_mad_f32 v{{[0-9]+}}, [[CVT_A]], [[CVT_B]], -[[CVT_C]]
279
280; VI-NOT: v_mac_f16
281; VI:     v_mad_f16 v{{[0-9]+}}, v{{[0-9]+}}, v{{[0-9]+}}, -v{{[0-9]}}
282; GCN:    s_endpgm
283define amdgpu_kernel void @mac_f16_neg_c_nsz_fp_math(
284    half addrspace(1)* %r,
285    half addrspace(1)* %a,
286    half addrspace(1)* %b,
287    half addrspace(1)* %c) #1 {
288entry:
289  %a.val = load half, half addrspace(1)* %a
290  %b.val = load half, half addrspace(1)* %b
291  %c.val = load half, half addrspace(1)* %c
292
293  %c.neg = fsub half 0.0, %c.val
294  %t.val = fmul half %a.val, %b.val
295  %r.val = fadd half %t.val, %c.neg
296
297  store half %r.val, half addrspace(1)* %r
298  ret void
299}
300
301; GCN-LABEL: {{^}}mac_v2f16:
302; GCN: {{buffer|flat}}_load_dword v[[A_V2_F16:[0-9]+]]
303; GCN: {{buffer|flat}}_load_dword v[[B_V2_F16:[0-9]+]]
304; GCN: {{buffer|flat}}_load_dword v[[C_V2_F16:[0-9]+]]
305
306; SI:  v_cvt_f32_f16_e32 v[[A_F32_0:[0-9]+]], v[[A_V2_F16]]
307; SI:  v_lshrrev_b32_e32 v[[A_F16_1:[0-9]+]], 16, v[[A_V2_F16]]
308; SI:  v_cvt_f32_f16_e32 v[[A_F32_1:[0-9]+]], v[[A_F16_1]]
309; SI:  v_lshrrev_b32_e32 v[[B_F16_1:[0-9]+]], 16, v[[B_V2_F16]]
310; SI:  v_cvt_f32_f16_e32 v[[B_F32_1:[0-9]+]], v[[B_F16_1]]
311; SI:  v_cvt_f32_f16_e32 v[[B_F32_0:[0-9]+]], v[[B_V2_F16]]
312; SI:  v_lshrrev_b32_e32 v[[C_F16_1:[0-9]+]], 16, v[[C_V2_F16]]
313; SI:  v_cvt_f32_f16_e32 v[[C_F32_1:[0-9]+]], v[[C_F16_1]]
314; SI-DAG:  v_cvt_f32_f16_e32 v[[C_F32_0:[0-9]+]], v[[C_V2_F16]]
315; SI-DAG:  v_mac_f32_e32 v[[C_F32_0]], v[[A_F32_0]], v[[B_F32_0]]
316; SI-DAG:  v_cvt_f16_f32_e32 v[[R_F16_LO:[0-9]+]], v[[C_F32_0]]
317; SI-DAG:  v_mac_f32_e32 v[[C_F32_1]], v[[A_F32_1]], v[[B_F32_1]]
318; SI-DAG:  v_cvt_f16_f32_e32 v[[R_F16_1:[0-9]+]], v[[C_F32_1]]
319; SI:  v_lshlrev_b32_e32 v[[R_F16_HI:[0-9]+]], 16, v[[R_F16_1]]
320; VI-NOT: and
321; SI:  v_or_b32_e32 v[[R_V2_F16:[0-9]+]], v[[R_F16_LO]], v[[R_F16_HI]]
322
323; VI-DAG: v_lshrrev_b32_e32 v[[C_F16_1:[0-9]+]], 16, v[[C_V2_F16]]
324; VI-DAG: v_mac_f16_sdwa v[[C_F16_1]], v[[A_V2_F16]], v[[B_V2_F16]] dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:WORD_1
325; VI-DAG: v_mac_f16_e32 v[[C_V2_F16]], v[[A_V2_F16]], v[[B_V2_F16]]
326; VI-DAG: v_lshlrev_b32_e32 v[[R_F16_HI:[0-9]+]], 16, v[[C_F16_1]]
327; VI-NOT: and
328; VI:  v_or_b32_e32 v[[R_V2_F16:[0-9]+]], v[[C_V2_F16]], v[[R_F16_HI]]
329
330; GCN: {{buffer|flat}}_store_dword v[[R_V2_F16]]
331; GCN: s_endpgm
332define amdgpu_kernel void @mac_v2f16(
333    <2 x half> addrspace(1)* %r,
334    <2 x half> addrspace(1)* %a,
335    <2 x half> addrspace(1)* %b,
336    <2 x half> addrspace(1)* %c) #0 {
337entry:
338  %a.val = load <2 x half>, <2 x half> addrspace(1)* %a
339  call void @llvm.amdgcn.s.barrier() #2
340  %b.val = load <2 x half>, <2 x half> addrspace(1)* %b
341  call void @llvm.amdgcn.s.barrier() #2
342  %c.val = load <2 x half>, <2 x half> addrspace(1)* %c
343
344  %t.val = fmul <2 x half> %a.val, %b.val
345  %r.val = fadd <2 x half> %t.val, %c.val
346
347  store <2 x half> %r.val, <2 x half> addrspace(1)* %r
348  ret void
349}
350
351; GCN-LABEL: {{^}}mac_v2f16_same_add:
352; SI:  v_mad_f32 v{{[0-9]}}, v{{[0-9]+}}, v{{[0-9]+}}, v{{[0-9]+}}
353; SI:  v_mac_f32_e32 v{{[0-9]+}}, v{{[0-9]+}}, v{{[0-9]+}}
354; SI:  v_mad_f32 v{{[0-9]}}, v{{[0-9]+}}, v{{[0-9]+}}, v{{[0-9]+}}
355; SI:  v_mac_f32_e32 v{{[0-9]+}}, v{{[0-9]+}}, v{{[0-9]+}}
356
357; VI-DAG:  v_mac_f16_sdwa v{{[0-9]}}, v{{[0-9]+}}, v{{[0-9]+}} dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:WORD_1
358; VI-DAG:  v_mad_f16 v{{[0-9]}}, v{{[0-9]+}}, v{{[0-9]+}}, v{{[0-9]+}}
359; VI-DAG:  v_mac_f16_sdwa v{{[0-9]+}}, v{{[0-9]+}}, v{{[0-9]+}} dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:WORD_1
360; VI-DAG:  v_mac_f16_e32 v{{[0-9]+}}, v{{[0-9]+}}, v{{[0-9]+}}
361
362; GCN: s_endpgm
363define amdgpu_kernel void @mac_v2f16_same_add(
364    <2 x half> addrspace(1)* %r0,
365    <2 x half> addrspace(1)* %r1,
366    <2 x half> addrspace(1)* %a,
367    <2 x half> addrspace(1)* %b,
368    <2 x half> addrspace(1)* %c,
369    <2 x half> addrspace(1)* %d,
370    <2 x half> addrspace(1)* %e) #0 {
371entry:
372  %a.val = load <2 x half>, <2 x half> addrspace(1)* %a
373  %b.val = load <2 x half>, <2 x half> addrspace(1)* %b
374  %c.val = load <2 x half>, <2 x half> addrspace(1)* %c
375  %d.val = load <2 x half>, <2 x half> addrspace(1)* %d
376  %e.val = load <2 x half>, <2 x half> addrspace(1)* %e
377
378  %t0.val = fmul <2 x half> %a.val, %b.val
379  %r0.val = fadd <2 x half> %t0.val, %c.val
380
381  %t1.val = fmul <2 x half> %d.val, %e.val
382  %r1.val = fadd <2 x half> %t1.val, %c.val
383
384  store <2 x half> %r0.val, <2 x half> addrspace(1)* %r0
385  store <2 x half> %r1.val, <2 x half> addrspace(1)* %r1
386  ret void
387}
388
389; GCN-LABEL: {{^}}mac_v2f16_neg_a:
390; SI: v_cvt_f32_f16_e32 [[CVT0:v[0-9]+]], {{v[0-9]+}}
391; SI: v_cvt_f32_f16_e32 [[CVT1:v[0-9]+]], {{v[0-9]+}}
392
393; SI: v_mad_f32 v{{[0-9]+}}, -v{{[0-9]+}}, v{{[0-9]+}}, v{{[0-9]+}}
394; SI: v_mad_f32 v{{[0-9]+}}, -v{{[0-9]+}}, v{{[0-9]+}}, v{{[0-9]+}}
395
396; VI-NOT: v_mac_f16
397; VI:     v_mad_f16 v{{[0-9]+}}, -v{{[0-9]+}}, v{{[0-9]+}}, v{{[0-9]+}}
398; VI:     v_mad_f16 v{{[0-9]+}}, -v{{[0-9]+}}, v{{[0-9]+}}, v{{[0-9]+}}
399; GCN:    s_endpgm
400define amdgpu_kernel void @mac_v2f16_neg_a(
401    <2 x half> addrspace(1)* %r,
402    <2 x half> addrspace(1)* %a,
403    <2 x half> addrspace(1)* %b,
404    <2 x half> addrspace(1)* %c) #0 {
405entry:
406  %a.val = load <2 x half>, <2 x half> addrspace(1)* %a
407  %b.val = load <2 x half>, <2 x half> addrspace(1)* %b
408  %c.val = load <2 x half>, <2 x half> addrspace(1)* %c
409
410  %a.neg = fsub <2 x half> <half -0.0, half -0.0>, %a.val
411  %t.val = fmul <2 x half> %a.neg, %b.val
412  %r.val = fadd <2 x half> %t.val, %c.val
413
414  store <2 x half> %r.val, <2 x half> addrspace(1)* %r
415  ret void
416}
417
418; GCN-LABEL: {{^}}mac_v2f16_neg_b
419; SI: v_cvt_f32_f16_e32 [[CVT0:v[0-9]+]], {{v[0-9]+}}
420; SI: v_cvt_f32_f16_e32 [[CVT1:v[0-9]+]], {{v[0-9]+}}
421; SI: v_mad_f32 v{{[0-9]+}}, -v{{[0-9]+}}, v{{[0-9]+}}, v{{[0-9]+}}
422; SI: v_mad_f32 v{{[0-9]+}}, -v{{[0-9]+}}, v{{[0-9]+}}, v{{[0-9]+}}
423
424
425; VI-NOT: v_mac_f16
426; VI:     v_mad_f16 v{{[0-9]+}}, -v{{[0-9]+}}, v{{[0-9]+}}, v{{[0-9]+}}
427; VI:     v_mad_f16 v{{[0-9]+}}, -v{{[0-9]+}}, v{{[0-9]+}}, v{{[0-9]+}}
428; GCN:    s_endpgm
429define amdgpu_kernel void @mac_v2f16_neg_b(
430    <2 x half> addrspace(1)* %r,
431    <2 x half> addrspace(1)* %a,
432    <2 x half> addrspace(1)* %b,
433    <2 x half> addrspace(1)* %c) #0 {
434entry:
435  %a.val = load <2 x half>, <2 x half> addrspace(1)* %a
436  %b.val = load <2 x half>, <2 x half> addrspace(1)* %b
437  %c.val = load <2 x half>, <2 x half> addrspace(1)* %c
438
439  %b.neg = fsub <2 x half> <half -0.0, half -0.0>, %b.val
440  %t.val = fmul <2 x half> %a.val, %b.neg
441  %r.val = fadd <2 x half> %t.val, %c.val
442
443  store <2 x half> %r.val, <2 x half> addrspace(1)* %r
444  ret void
445}
446
447; GCN-LABEL: {{^}}mac_v2f16_neg_c:
448; SI: v_cvt_f32_f16_e32 [[CVT0:v[0-9]+]], {{v[0-9]+}}
449; SI: v_cvt_f32_f16_e32 [[CVT1:v[0-9]+]], {{v[0-9]+}}
450; SI: v_cvt_f32_f16_e32 [[CVT2:v[0-9]+]], {{v[0-9]+}}
451; SI: v_cvt_f32_f16_e32 [[CVT3:v[0-9]+]], {{v[0-9]+}}
452; SI: v_cvt_f32_f16_e32 [[CVT4:v[0-9]+]], {{v[0-9]+}}
453; SI: v_cvt_f32_f16_e32 [[CVT5:v[0-9]+]], {{v[0-9]+}}
454
455; SI-DAG: v_mad_f32 v{{[0-9]+}}, v{{[0-9]+}}, v{{[0-9]+}}, -v{{[0-9]+}}
456; SI-DAG: v_mad_f32 v{{[0-9]+}}, v{{[0-9]+}}, v{{[0-9]+}}, -v{{[0-9]+}}
457
458; VI-NOT: v_mac_f16
459; VI:     v_mad_f16 v{{[0-9]+}}, v{{[0-9]+}}, v{{[0-9]+}}, -v{{[0-9]+}}
460; VI:     v_mad_f16 v{{[0-9]+}}, v{{[0-9]+}}, v{{[0-9]+}}, -v{{[0-9]+}}
461; GCN:    s_endpgm
462define amdgpu_kernel void @mac_v2f16_neg_c(
463    <2 x half> addrspace(1)* %r,
464    <2 x half> addrspace(1)* %a,
465    <2 x half> addrspace(1)* %b,
466    <2 x half> addrspace(1)* %c) #0 {
467entry:
468  %a.val = load <2 x half>, <2 x half> addrspace(1)* %a
469  %b.val = load <2 x half>, <2 x half> addrspace(1)* %b
470  %c.val = load <2 x half>, <2 x half> addrspace(1)* %c
471
472  %c.neg = fsub <2 x half> <half -0.0, half -0.0>, %c.val
473  %t.val = fmul <2 x half> %a.val, %b.val
474  %r.val = fadd <2 x half> %t.val, %c.neg
475
476  store <2 x half> %r.val, <2 x half> addrspace(1)* %r
477  ret void
478}
479
480; GCN-LABEL: {{^}}mac_v2f16_neg_a_safe_fp_math:
481
482; SI:  v_sub_f32_e32 v[[NEG_A0:[0-9]+]], 0, v{{[0-9]+}}
483; SI:  v_sub_f32_e32 v[[NEG_A1:[0-9]+]], 0, v{{[0-9]+}}
484; SI-DAG:  v_mac_f32_e32 v{{[0-9]+}}, v[[NEG_A0]], v{{[0-9]+}}
485; SI-DAG:  v_mac_f32_e32 v{{[0-9]+}}, v[[NEG_A1]], v{{[0-9]+}}
486
487; VI-DAG:  v_mov_b32_e32 [[ZERO:v[0-9]+]], 0
488; VI-DAG:  v_sub_f16_e32 v[[NEG_A1:[0-9]+]], 0, v{{[0-9]+}}
489; VI-DAG:  v_sub_f16_sdwa v[[NEG_A0:[0-9]+]], [[ZERO]], v{{[0-9]+}} dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_1
490; VI-DAG:  v_mac_f16_sdwa v{{[0-9]+}}, v[[NEG_A0]], v{{[0-9]+}} dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_1
491; VI-DAG:  v_mac_f16_e32 v{{[0-9]+}}, v[[NEG_A1]], v{{[0-9]+}}
492
493; GCN: s_endpgm
494define amdgpu_kernel void @mac_v2f16_neg_a_safe_fp_math(
495    <2 x half> addrspace(1)* %r,
496    <2 x half> addrspace(1)* %a,
497    <2 x half> addrspace(1)* %b,
498    <2 x half> addrspace(1)* %c) #0 {
499entry:
500  %a.val = load <2 x half>, <2 x half> addrspace(1)* %a
501  %b.val = load <2 x half>, <2 x half> addrspace(1)* %b
502  %c.val = load <2 x half>, <2 x half> addrspace(1)* %c
503
504  %a.neg = fsub <2 x half> <half 0.0, half 0.0>, %a.val
505  %t.val = fmul <2 x half> %a.neg, %b.val
506  %r.val = fadd <2 x half> %t.val, %c.val
507
508  store <2 x half> %r.val, <2 x half> addrspace(1)* %r
509  ret void
510}
511
512; GCN-LABEL: {{^}}mac_v2f16_neg_b_safe_fp_math:
513
514; SI:  v_sub_f32_e32 v[[NEG_A0:[0-9]+]], 0, v{{[0-9]+}}
515; SI:  v_sub_f32_e32 v[[NEG_A1:[0-9]+]], 0, v{{[0-9]+}}
516; SI-DAG:  v_mac_f32_e32 v{{[0-9]+}}, v{{[0-9]+}}, v[[NEG_A0]]
517; SI-DAG:  v_mac_f32_e32 v{{[0-9]+}}, v{{[0-9]+}}, v[[NEG_A1]]
518
519; VI:  v_mov_b32_e32 [[ZERO:v[0-9]+]], 0
520; VI:  v_sub_f16_e32 v[[NEG_A1:[0-9]+]], 0, v{{[0-9]+}}
521; VI:  v_sub_f16_sdwa v[[NEG_A0:[0-9]+]], [[ZERO]], v{{[0-9]+}} dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_1
522; VI-DAG:  v_mac_f16_sdwa v{{[0-9]+}}, v{{[0-9]+}}, v[[NEG_A0]] dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
523; VI-DAG:  v_mac_f16_e32 v{{[0-9]+}}, v{{[0-9]+}}, v[[NEG_A1]]
524
525; GCN: s_endpgm
526define amdgpu_kernel void @mac_v2f16_neg_b_safe_fp_math(
527    <2 x half> addrspace(1)* %r,
528    <2 x half> addrspace(1)* %a,
529    <2 x half> addrspace(1)* %b,
530    <2 x half> addrspace(1)* %c) #0 {
531entry:
532  %a.val = load <2 x half>, <2 x half> addrspace(1)* %a
533  %b.val = load <2 x half>, <2 x half> addrspace(1)* %b
534  %c.val = load <2 x half>, <2 x half> addrspace(1)* %c
535
536  %b.neg = fsub <2 x half> <half 0.0, half 0.0>, %b.val
537  %t.val = fmul <2 x half> %a.val, %b.neg
538  %r.val = fadd <2 x half> %t.val, %c.val
539
540  store <2 x half> %r.val, <2 x half> addrspace(1)* %r
541  ret void
542}
543
544; GCN-LABEL: {{^}}mac_v2f16_neg_c_safe_fp_math:
545
546; SI:  v_sub_f32_e32 v[[NEG_A0:[0-9]+]], 0, v{{[0-9]+}}
547; SI:  v_sub_f32_e32 v[[NEG_A1:[0-9]+]], 0, v{{[0-9]+}}
548; SI-DAG:  v_mac_f32_e32 v[[NEG_A0]], v{{[0-9]+}}, v{{[0-9]+}}
549; SI-DAG:  v_mac_f32_e32 v[[NEG_A1]], v{{[0-9]+}}, v{{[0-9]+}}
550
551; VI:  v_mov_b32_e32 [[ZERO:v[0-9]+]], 0
552; VI:  v_sub_f16_e32 v[[NEG_A1:[0-9]+]], 0, v{{[0-9]+}}
553; VI:  v_sub_f16_sdwa v[[NEG_A0:[0-9]+]], [[ZERO]], v{{[0-9]+}} dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_1
554; VI-DAG:  v_mac_f16_sdwa v[[NEG_A0]], v{{[0-9]+}}, v{{[0-9]+}} dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:WORD_1
555; VI-DAG:  v_mac_f16_e32 v[[NEG_A1]], v{{[0-9]+}}, v{{[0-9]+}}
556
557; GCN: s_endpgm
558define amdgpu_kernel void @mac_v2f16_neg_c_safe_fp_math(
559    <2 x half> addrspace(1)* %r,
560    <2 x half> addrspace(1)* %a,
561    <2 x half> addrspace(1)* %b,
562    <2 x half> addrspace(1)* %c) #0 {
563entry:
564  %a.val = load <2 x half>, <2 x half> addrspace(1)* %a
565  %b.val = load <2 x half>, <2 x half> addrspace(1)* %b
566  %c.val = load <2 x half>, <2 x half> addrspace(1)* %c
567
568  %c.neg = fsub <2 x half> <half 0.0, half 0.0>, %c.val
569  %t.val = fmul <2 x half> %a.val, %b.val
570  %r.val = fadd <2 x half> %t.val, %c.neg
571
572  store <2 x half> %r.val, <2 x half> addrspace(1)* %r
573  ret void
574}
575
576; GCN-LABEL: {{^}}mac_v2f16_neg_a_nsz_fp_math:
577; SI: v_cvt_f32_f16_e32 [[CVT0:v[0-9]+]], {{v[0-9]+}}
578; SI: v_cvt_f32_f16_e32 [[CVT1:v[0-9]+]], {{v[0-9]+}}
579; SI: v_cvt_f32_f16_e32 [[CVT2:v[0-9]+]], {{v[0-9]+}}
580; SI: v_cvt_f32_f16_e32 [[CVT3:v[0-9]+]], {{v[0-9]+}}
581; SI: v_cvt_f32_f16_e32 [[CVT4:v[0-9]+]], {{v[0-9]+}}
582; SI: v_cvt_f32_f16_e32 [[CVT5:v[0-9]+]], {{v[0-9]+}}
583
584; SI-DAG: v_mad_f32 v{{[0-9]+}}, -v{{[0-9]+}}, v{{[0-9]+}}, v{{[0-9]+}}
585; SI-DAG: v_mad_f32 v{{[0-9]+}}, -v{{[0-9]+}}, v{{[0-9]+}}, v{{[0-9]+}}
586
587; VI-NOT: v_mac_f16
588; VI:     v_mad_f16 v{{[0-9]+}}, -v{{[0-9]+}}, v{{[0-9]+}}, v{{[-0-9]}}
589; VI:     v_mad_f16 v{{[0-9]+}}, -v{{[0-9]+}}, v{{[0-9]+}}, v{{[-0-9]}}
590; GCN:    s_endpgm
591define amdgpu_kernel void @mac_v2f16_neg_a_nsz_fp_math(
592    <2 x half> addrspace(1)* %r,
593    <2 x half> addrspace(1)* %a,
594    <2 x half> addrspace(1)* %b,
595    <2 x half> addrspace(1)* %c) #1 {
596entry:
597  %a.val = load <2 x half>, <2 x half> addrspace(1)* %a
598  %b.val = load <2 x half>, <2 x half> addrspace(1)* %b
599  %c.val = load <2 x half>, <2 x half> addrspace(1)* %c
600
601  %a.neg = fsub <2 x half> <half 0.0, half 0.0>, %a.val
602  %t.val = fmul <2 x half> %a.neg, %b.val
603  %r.val = fadd <2 x half> %t.val, %c.val
604
605  store <2 x half> %r.val, <2 x half> addrspace(1)* %r
606  ret void
607}
608
609; GCN-LABEL: {{^}}mac_v2f16_neg_b_nsz_fp_math:
610; SI: v_cvt_f32_f16_e32 [[CVT0:v[0-9]+]], {{v[0-9]+}}
611; SI: v_cvt_f32_f16_e32 [[CVT1:v[0-9]+]], {{v[0-9]+}}
612; SI: v_cvt_f32_f16_e32 [[CVT2:v[0-9]+]], {{v[0-9]+}}
613; SI: v_cvt_f32_f16_e32 [[CVT3:v[0-9]+]], {{v[0-9]+}}
614; SI: v_cvt_f32_f16_e32 [[CVT4:v[0-9]+]], {{v[0-9]+}}
615; SI: v_cvt_f32_f16_e32 [[CVT5:v[0-9]+]], {{v[0-9]+}}
616
617; SI-DAG: v_mad_f32 v{{[0-9]+}}, -v{{[0-9]+}}, v{{[0-9]+}}, v{{[0-9]+}}
618; SI-DAG: v_mad_f32 v{{[0-9]+}}, -v{{[0-9]+}}, v{{[0-9]+}}, v{{[0-9]+}}
619
620; VI-NOT: v_mac_f16
621; VI:     v_mad_f16 v{{[0-9]+}}, -v{{[0-9]+}}, v{{[0-9]+}}, v{{[-0-9]}}
622; VI:     v_mad_f16 v{{[0-9]+}}, -v{{[0-9]+}}, v{{[0-9]+}}, v{{[-0-9]}}
623; GCN:    s_endpgm
624define amdgpu_kernel void @mac_v2f16_neg_b_nsz_fp_math(
625    <2 x half> addrspace(1)* %r,
626    <2 x half> addrspace(1)* %a,
627    <2 x half> addrspace(1)* %b,
628    <2 x half> addrspace(1)* %c) #1 {
629entry:
630  %a.val = load <2 x half>, <2 x half> addrspace(1)* %a
631  %b.val = load <2 x half>, <2 x half> addrspace(1)* %b
632  %c.val = load <2 x half>, <2 x half> addrspace(1)* %c
633
634  %b.neg = fsub <2 x half> <half 0.0, half 0.0>, %b.val
635  %t.val = fmul <2 x half> %a.val, %b.neg
636  %r.val = fadd <2 x half> %t.val, %c.val
637
638  store <2 x half> %r.val, <2 x half> addrspace(1)* %r
639  ret void
640}
641
642; GCN-LABEL: {{^}}mac_v2f16_neg_c_nsz_fp_math:
643; SI: v_cvt_f32_f16_e32 [[CVT0:v[0-9]+]], {{v[0-9]+}}
644; SI: v_cvt_f32_f16_e32 [[CVT1:v[0-9]+]], {{v[0-9]+}}
645; SI: v_cvt_f32_f16_e32 [[CVT2:v[0-9]+]], {{v[0-9]+}}
646; SI: v_cvt_f32_f16_e32 [[CVT3:v[0-9]+]], {{v[0-9]+}}
647; SI: v_cvt_f32_f16_e32 [[CVT4:v[0-9]+]], {{v[0-9]+}}
648; SI: v_cvt_f32_f16_e32 [[CVT5:v[0-9]+]], {{v[0-9]+}}
649
650; SI-DAG: v_mad_f32 v{{[0-9]+}}, v{{[0-9]+}}, v{{[0-9]+}}, -v{{[0-9]+}}
651; SI-DAG: v_mad_f32 v{{[0-9]+}}, v{{[0-9]+}}, v{{[0-9]+}}, -v{{[0-9]+}}
652
653; VI-NOT: v_mac_f16
654; VI:     v_mad_f16 v{{[0-9]+}}, v{{[0-9]+}}, v{{[0-9]+}}, -v{{[-0-9]}}
655; VI:     v_mad_f16 v{{[0-9]+}}, v{{[0-9]+}}, v{{[0-9]+}}, -v{{[-0-9]}}
656; GCN:    s_endpgm
657define amdgpu_kernel void @mac_v2f16_neg_c_nsz_fp_math(
658    <2 x half> addrspace(1)* %r,
659    <2 x half> addrspace(1)* %a,
660    <2 x half> addrspace(1)* %b,
661    <2 x half> addrspace(1)* %c) #1 {
662entry:
663  %a.val = load <2 x half>, <2 x half> addrspace(1)* %a
664  %b.val = load <2 x half>, <2 x half> addrspace(1)* %b
665  %c.val = load <2 x half>, <2 x half> addrspace(1)* %c
666
667  %c.neg = fsub <2 x half> <half 0.0, half 0.0>, %c.val
668  %t.val = fmul <2 x half> %a.val, %b.val
669  %r.val = fadd <2 x half> %t.val, %c.neg
670
671  store <2 x half> %r.val, <2 x half> addrspace(1)* %r
672  ret void
673}
674
675declare void @llvm.amdgcn.s.barrier() #2
676
677attributes #0 = { nounwind "no-signed-zeros-fp-math"="false" }
678attributes #1 = { nounwind "no-signed-zeros-fp-math"="true" }
679attributes #2 = { nounwind convergent }
680