[llvm-project.git] / llvm / test / Transforms / AtomicExpand / AMDGPU / expand-atomic-v2f16-system.ll
| blob | 19f02dea21329711cbae9ae186290bcab628c933 |
1 ; NOTE: Assertions have been autogenerated by utils/update_test_checks.py UTC_ARGS: --version 4
2 ; RUN: opt -S -mtriple=amdgcn-amd-amdhsa -mcpu=gfx803 -passes=atomic-expand %s | FileCheck -check-prefixes=COMMON,GFX803 %s
3 ; RUN: opt -S -mtriple=amdgcn-amd-amdhsa -mcpu=gfx906 -passes=atomic-expand %s | FileCheck -check-prefixes=COMMON,GFX906 %s
4 ; RUN: opt -S -mtriple=amdgcn-amd-amdhsa -mcpu=gfx908 -passes=atomic-expand %s | FileCheck -check-prefixes=COMMON,GFX908 %s
5 ; RUN: opt -S -mtriple=amdgcn-amd-amdhsa -mcpu=gfx90a -passes=atomic-expand %s | FileCheck -check-prefixes=COMMON,GFX90A %s
6 ; RUN: opt -S -mtriple=amdgcn-amd-amdhsa -mcpu=gfx940 -passes=atomic-expand %s | FileCheck -check-prefixes=COMMON,GFX940 %s
7 ; RUN: opt -S -mtriple=amdgcn-amd-amdhsa -mcpu=gfx1030 -passes=atomic-expand %s | FileCheck -check-prefixes=COMMON,GFX10 %s
8 ; RUN: opt -S -mtriple=amdgcn-amd-amdhsa -mcpu=gfx1100 -passes=atomic-expand %s | FileCheck -check-prefixes=COMMON,GFX11 %s
9 ; RUN: opt -S -mtriple=amdgcn-amd-amdhsa -mcpu=gfx1200 -passes=atomic-expand %s | FileCheck -check-prefixes=COMMON,GFX12 %s
11 ;---------------------------------------------------------------------
12 ; TODO: atomicrmw xchg
13 ;---------------------------------------------------------------------
15 ; ; xchg is supported over PCIe, so no expansion is necessary
16 ; define <2 x half> @test_atomicrmw_xchg_v2f16_global_system(ptr addrspace(1) %ptr, <2 x half> %value) {
17 ; %res = atomicrmw xchg ptr addrspace(1) %ptr, <2 x half> %value seq_cst
18 ; ret <2 x half> %res
19 ; }
21 ; ; xchg is supported over PCIe, so no expansion is necessary. Metadata should be ignored.
22 ; define <2 x half> @test_atomicrmw_xchg_v2f16_global_system__amdgpu_no_fine_grained_memory(ptr addrspace(1) %ptr, <2 x half> %value) {
23 ; %res = atomicrmw xchg ptr addrspace(1) %ptr, <2 x half> %value seq_cst, !amdgpu.no.fine.grained.memory !0
24 ; ret <2 x half> %res
25 ; }
27 ; ; xchg is supported over PCIe, so no expansion is necessary. Metadata should be ignored.
28 ; define <2 x half> @test_atomicrmw_xchg_v2f16_global_system__amdgpu_no_remote_memory(ptr addrspace(1) %ptr, <2 x half> %value) {
29 ; %res = atomicrmw xchg ptr addrspace(1) %ptr, <2 x half> %value seq_cst, !amdgpu.no.remote.memory !0
30 ; ret <2 x half> %res
31 ; }
33 ; ; xchg is supported over PCIe, so no expansion is necessary. Metadata should be ignored.
34 ; define <2 x half> @test_atomicrmw_xchg_v2f16_global_system__amdgpu_no_fine_grained_memory__amdgpu_no_remote_memory(ptr addrspace(1) %ptr, <2 x half> %value) {
35 ; %res = atomicrmw xchg ptr addrspace(1) %ptr, <2 x half> %value seq_cst, !amdgpu.no.fine.grained.memory !0, !amdgpu.no.remote.memory !0
36 ; ret <2 x half> %res
37 ; }
39 ;---------------------------------------------------------------------
40 ; atomicrmw fadd
41 ;---------------------------------------------------------------------
43 define <2 x half> @test_atomicrmw_fadd_v2f16_global_system(ptr addrspace(1) %ptr, <2 x half> %value) {
44 ; COMMON-LABEL: define <2 x half> @test_atomicrmw_fadd_v2f16_global_system(
45 ; COMMON-SAME: ptr addrspace(1) [[PTR:%.*]], <2 x half> [[VALUE:%.*]]) #[[ATTR0:[0-9]+]] {
46 ; COMMON-NEXT: [[TMP1:%.*]] = load <2 x half>, ptr addrspace(1) [[PTR]], align 4
47 ; COMMON-NEXT: br label [[ATOMICRMW_START:%.*]]
48 ; COMMON: atomicrmw.start:
49 ; COMMON-NEXT: [[LOADED:%.*]] = phi <2 x half> [ [[TMP1]], [[TMP0:%.*]] ], [ [[TMP5:%.*]], [[ATOMICRMW_START]] ]
50 ; COMMON-NEXT: [[NEW:%.*]] = fadd <2 x half> [[LOADED]], [[VALUE]]
51 ; COMMON-NEXT: [[TMP2:%.*]] = bitcast <2 x half> [[NEW]] to i32
52 ; COMMON-NEXT: [[TMP3:%.*]] = bitcast <2 x half> [[LOADED]] to i32
53 ; COMMON-NEXT: [[TMP4:%.*]] = cmpxchg ptr addrspace(1) [[PTR]], i32 [[TMP3]], i32 [[TMP2]] seq_cst seq_cst, align 4
54 ; COMMON-NEXT: [[SUCCESS:%.*]] = extractvalue { i32, i1 } [[TMP4]], 1
55 ; COMMON-NEXT: [[NEWLOADED:%.*]] = extractvalue { i32, i1 } [[TMP4]], 0
56 ; COMMON-NEXT: [[TMP5]] = bitcast i32 [[NEWLOADED]] to <2 x half>
57 ; COMMON-NEXT: br i1 [[SUCCESS]], label [[ATOMICRMW_END:%.*]], label [[ATOMICRMW_START]]
58 ; COMMON: atomicrmw.end:
59 ; COMMON-NEXT: ret <2 x half> [[TMP5]]
60 ;
61 %res = atomicrmw fadd ptr addrspace(1) %ptr, <2 x half> %value seq_cst
62 ret <2 x half> %res
63 }
65 define <2 x half> @test_atomicrmw_fadd_v2f16_global_system__amdgpu_no_fine_grained_memory(ptr addrspace(1) %ptr, <2 x half> %value) {
66 ; GFX803-LABEL: define <2 x half> @test_atomicrmw_fadd_v2f16_global_system__amdgpu_no_fine_grained_memory(
67 ; GFX803-SAME: ptr addrspace(1) [[PTR:%.*]], <2 x half> [[VALUE:%.*]]) #[[ATTR0]] {
68 ; GFX803-NEXT: [[TMP1:%.*]] = load <2 x half>, ptr addrspace(1) [[PTR]], align 4
69 ; GFX803-NEXT: br label [[ATOMICRMW_START:%.*]]
70 ; GFX803: atomicrmw.start:
71 ; GFX803-NEXT: [[LOADED:%.*]] = phi <2 x half> [ [[TMP1]], [[TMP0:%.*]] ], [ [[TMP5:%.*]], [[ATOMICRMW_START]] ]
72 ; GFX803-NEXT: [[NEW:%.*]] = fadd <2 x half> [[LOADED]], [[VALUE]]
73 ; GFX803-NEXT: [[TMP2:%.*]] = bitcast <2 x half> [[NEW]] to i32
74 ; GFX803-NEXT: [[TMP3:%.*]] = bitcast <2 x half> [[LOADED]] to i32
75 ; GFX803-NEXT: [[TMP4:%.*]] = cmpxchg ptr addrspace(1) [[PTR]], i32 [[TMP3]], i32 [[TMP2]] seq_cst seq_cst, align 4, !amdgpu.no.fine.grained.memory [[META0:![0-9]+]]
76 ; GFX803-NEXT: [[SUCCESS:%.*]] = extractvalue { i32, i1 } [[TMP4]], 1
77 ; GFX803-NEXT: [[NEWLOADED:%.*]] = extractvalue { i32, i1 } [[TMP4]], 0
78 ; GFX803-NEXT: [[TMP5]] = bitcast i32 [[NEWLOADED]] to <2 x half>
79 ; GFX803-NEXT: br i1 [[SUCCESS]], label [[ATOMICRMW_END:%.*]], label [[ATOMICRMW_START]]
80 ; GFX803: atomicrmw.end:
81 ; GFX803-NEXT: ret <2 x half> [[TMP5]]
82 ;
83 ; GFX906-LABEL: define <2 x half> @test_atomicrmw_fadd_v2f16_global_system__amdgpu_no_fine_grained_memory(
84 ; GFX906-SAME: ptr addrspace(1) [[PTR:%.*]], <2 x half> [[VALUE:%.*]]) #[[ATTR0]] {
85 ; GFX906-NEXT: [[TMP1:%.*]] = load <2 x half>, ptr addrspace(1) [[PTR]], align 4
86 ; GFX906-NEXT: br label [[ATOMICRMW_START:%.*]]
87 ; GFX906: atomicrmw.start:
88 ; GFX906-NEXT: [[LOADED:%.*]] = phi <2 x half> [ [[TMP1]], [[TMP0:%.*]] ], [ [[TMP5:%.*]], [[ATOMICRMW_START]] ]
89 ; GFX906-NEXT: [[NEW:%.*]] = fadd <2 x half> [[LOADED]], [[VALUE]]
90 ; GFX906-NEXT: [[TMP2:%.*]] = bitcast <2 x half> [[NEW]] to i32
91 ; GFX906-NEXT: [[TMP3:%.*]] = bitcast <2 x half> [[LOADED]] to i32
92 ; GFX906-NEXT: [[TMP4:%.*]] = cmpxchg ptr addrspace(1) [[PTR]], i32 [[TMP3]], i32 [[TMP2]] seq_cst seq_cst, align 4, !amdgpu.no.fine.grained.memory [[META0:![0-9]+]]
93 ; GFX906-NEXT: [[SUCCESS:%.*]] = extractvalue { i32, i1 } [[TMP4]], 1
94 ; GFX906-NEXT: [[NEWLOADED:%.*]] = extractvalue { i32, i1 } [[TMP4]], 0
95 ; GFX906-NEXT: [[TMP5]] = bitcast i32 [[NEWLOADED]] to <2 x half>
96 ; GFX906-NEXT: br i1 [[SUCCESS]], label [[ATOMICRMW_END:%.*]], label [[ATOMICRMW_START]]
97 ; GFX906: atomicrmw.end:
98 ; GFX906-NEXT: ret <2 x half> [[TMP5]]
99 ;
100 ; GFX908-LABEL: define <2 x half> @test_atomicrmw_fadd_v2f16_global_system__amdgpu_no_fine_grained_memory(
101 ; GFX908-SAME: ptr addrspace(1) [[PTR:%.*]], <2 x half> [[VALUE:%.*]]) #[[ATTR0]] {
102 ; GFX908-NEXT: [[TMP1:%.*]] = load <2 x half>, ptr addrspace(1) [[PTR]], align 4
103 ; GFX908-NEXT: br label [[ATOMICRMW_START:%.*]]
104 ; GFX908: atomicrmw.start:
105 ; GFX908-NEXT: [[LOADED:%.*]] = phi <2 x half> [ [[TMP1]], [[TMP0:%.*]] ], [ [[TMP5:%.*]], [[ATOMICRMW_START]] ]
106 ; GFX908-NEXT: [[NEW:%.*]] = fadd <2 x half> [[LOADED]], [[VALUE]]
107 ; GFX908-NEXT: [[TMP2:%.*]] = bitcast <2 x half> [[NEW]] to i32
108 ; GFX908-NEXT: [[TMP3:%.*]] = bitcast <2 x half> [[LOADED]] to i32
109 ; GFX908-NEXT: [[TMP4:%.*]] = cmpxchg ptr addrspace(1) [[PTR]], i32 [[TMP3]], i32 [[TMP2]] seq_cst seq_cst, align 4, !amdgpu.no.fine.grained.memory [[META0:![0-9]+]]
110 ; GFX908-NEXT: [[SUCCESS:%.*]] = extractvalue { i32, i1 } [[TMP4]], 1
111 ; GFX908-NEXT: [[NEWLOADED:%.*]] = extractvalue { i32, i1 } [[TMP4]], 0
112 ; GFX908-NEXT: [[TMP5]] = bitcast i32 [[NEWLOADED]] to <2 x half>
113 ; GFX908-NEXT: br i1 [[SUCCESS]], label [[ATOMICRMW_END:%.*]], label [[ATOMICRMW_START]]
114 ; GFX908: atomicrmw.end:
115 ; GFX908-NEXT: ret <2 x half> [[TMP5]]
116 ;
117 ; GFX90A-LABEL: define <2 x half> @test_atomicrmw_fadd_v2f16_global_system__amdgpu_no_fine_grained_memory(
118 ; GFX90A-SAME: ptr addrspace(1) [[PTR:%.*]], <2 x half> [[VALUE:%.*]]) #[[ATTR0]] {
119 ; GFX90A-NEXT: [[RES:%.*]] = atomicrmw fadd ptr addrspace(1) [[PTR]], <2 x half> [[VALUE]] seq_cst, align 4, !amdgpu.no.fine.grained.memory [[META0:![0-9]+]]
120 ; GFX90A-NEXT: ret <2 x half> [[RES]]
121 ;
122 ; GFX940-LABEL: define <2 x half> @test_atomicrmw_fadd_v2f16_global_system__amdgpu_no_fine_grained_memory(
123 ; GFX940-SAME: ptr addrspace(1) [[PTR:%.*]], <2 x half> [[VALUE:%.*]]) #[[ATTR0]] {
124 ; GFX940-NEXT: [[RES:%.*]] = atomicrmw fadd ptr addrspace(1) [[PTR]], <2 x half> [[VALUE]] seq_cst, align 4, !amdgpu.no.fine.grained.memory [[META0:![0-9]+]]
125 ; GFX940-NEXT: ret <2 x half> [[RES]]
126 ;
127 ; GFX10-LABEL: define <2 x half> @test_atomicrmw_fadd_v2f16_global_system__amdgpu_no_fine_grained_memory(
128 ; GFX10-SAME: ptr addrspace(1) [[PTR:%.*]], <2 x half> [[VALUE:%.*]]) #[[ATTR0]] {
129 ; GFX10-NEXT: [[TMP1:%.*]] = load <2 x half>, ptr addrspace(1) [[PTR]], align 4
130 ; GFX10-NEXT: br label [[ATOMICRMW_START:%.*]]
131 ; GFX10: atomicrmw.start:
132 ; GFX10-NEXT: [[LOADED:%.*]] = phi <2 x half> [ [[TMP1]], [[TMP0:%.*]] ], [ [[TMP5:%.*]], [[ATOMICRMW_START]] ]
133 ; GFX10-NEXT: [[NEW:%.*]] = fadd <2 x half> [[LOADED]], [[VALUE]]
134 ; GFX10-NEXT: [[TMP2:%.*]] = bitcast <2 x half> [[NEW]] to i32
135 ; GFX10-NEXT: [[TMP3:%.*]] = bitcast <2 x half> [[LOADED]] to i32
136 ; GFX10-NEXT: [[TMP4:%.*]] = cmpxchg ptr addrspace(1) [[PTR]], i32 [[TMP3]], i32 [[TMP2]] seq_cst seq_cst, align 4, !amdgpu.no.fine.grained.memory [[META0:![0-9]+]]
137 ; GFX10-NEXT: [[SUCCESS:%.*]] = extractvalue { i32, i1 } [[TMP4]], 1
138 ; GFX10-NEXT: [[NEWLOADED:%.*]] = extractvalue { i32, i1 } [[TMP4]], 0
139 ; GFX10-NEXT: [[TMP5]] = bitcast i32 [[NEWLOADED]] to <2 x half>
140 ; GFX10-NEXT: br i1 [[SUCCESS]], label [[ATOMICRMW_END:%.*]], label [[ATOMICRMW_START]]
141 ; GFX10: atomicrmw.end:
142 ; GFX10-NEXT: ret <2 x half> [[TMP5]]
143 ;
144 ; GFX11-LABEL: define <2 x half> @test_atomicrmw_fadd_v2f16_global_system__amdgpu_no_fine_grained_memory(
145 ; GFX11-SAME: ptr addrspace(1) [[PTR:%.*]], <2 x half> [[VALUE:%.*]]) #[[ATTR0]] {
146 ; GFX11-NEXT: [[TMP1:%.*]] = load <2 x half>, ptr addrspace(1) [[PTR]], align 4
147 ; GFX11-NEXT: br label [[ATOMICRMW_START:%.*]]
148 ; GFX11: atomicrmw.start:
149 ; GFX11-NEXT: [[LOADED:%.*]] = phi <2 x half> [ [[TMP1]], [[TMP0:%.*]] ], [ [[TMP5:%.*]], [[ATOMICRMW_START]] ]
150 ; GFX11-NEXT: [[NEW:%.*]] = fadd <2 x half> [[LOADED]], [[VALUE]]
151 ; GFX11-NEXT: [[TMP2:%.*]] = bitcast <2 x half> [[NEW]] to i32
152 ; GFX11-NEXT: [[TMP3:%.*]] = bitcast <2 x half> [[LOADED]] to i32
153 ; GFX11-NEXT: [[TMP4:%.*]] = cmpxchg ptr addrspace(1) [[PTR]], i32 [[TMP3]], i32 [[TMP2]] seq_cst seq_cst, align 4, !amdgpu.no.fine.grained.memory [[META0:![0-9]+]]
154 ; GFX11-NEXT: [[SUCCESS:%.*]] = extractvalue { i32, i1 } [[TMP4]], 1
155 ; GFX11-NEXT: [[NEWLOADED:%.*]] = extractvalue { i32, i1 } [[TMP4]], 0
156 ; GFX11-NEXT: [[TMP5]] = bitcast i32 [[NEWLOADED]] to <2 x half>
157 ; GFX11-NEXT: br i1 [[SUCCESS]], label [[ATOMICRMW_END:%.*]], label [[ATOMICRMW_START]]
158 ; GFX11: atomicrmw.end:
159 ; GFX11-NEXT: ret <2 x half> [[TMP5]]
160 ;
161 ; GFX12-LABEL: define <2 x half> @test_atomicrmw_fadd_v2f16_global_system__amdgpu_no_fine_grained_memory(
162 ; GFX12-SAME: ptr addrspace(1) [[PTR:%.*]], <2 x half> [[VALUE:%.*]]) #[[ATTR0]] {
163 ; GFX12-NEXT: [[RES:%.*]] = atomicrmw fadd ptr addrspace(1) [[PTR]], <2 x half> [[VALUE]] seq_cst, align 4, !amdgpu.no.fine.grained.memory [[META0:![0-9]+]]
164 ; GFX12-NEXT: ret <2 x half> [[RES]]
165 ;
166 %res = atomicrmw fadd ptr addrspace(1) %ptr, <2 x half> %value seq_cst, !amdgpu.no.fine.grained.memory !0
167 ret <2 x half> %res
168 }
170 define <2 x half> @test_atomicrmw_fadd_v2f16_global_system__amdgpu_no_remote_memory(ptr addrspace(1) %ptr, <2 x half> %value) {
171 ; GFX803-LABEL: define <2 x half> @test_atomicrmw_fadd_v2f16_global_system__amdgpu_no_remote_memory(
172 ; GFX803-SAME: ptr addrspace(1) [[PTR:%.*]], <2 x half> [[VALUE:%.*]]) #[[ATTR0]] {
173 ; GFX803-NEXT: [[TMP1:%.*]] = load <2 x half>, ptr addrspace(1) [[PTR]], align 4
174 ; GFX803-NEXT: br label [[ATOMICRMW_START:%.*]]
175 ; GFX803: atomicrmw.start:
176 ; GFX803-NEXT: [[LOADED:%.*]] = phi <2 x half> [ [[TMP1]], [[TMP0:%.*]] ], [ [[TMP5:%.*]], [[ATOMICRMW_START]] ]
177 ; GFX803-NEXT: [[NEW:%.*]] = fadd <2 x half> [[LOADED]], [[VALUE]]
178 ; GFX803-NEXT: [[TMP2:%.*]] = bitcast <2 x half> [[NEW]] to i32
179 ; GFX803-NEXT: [[TMP3:%.*]] = bitcast <2 x half> [[LOADED]] to i32
180 ; GFX803-NEXT: [[TMP4:%.*]] = cmpxchg ptr addrspace(1) [[PTR]], i32 [[TMP3]], i32 [[TMP2]] seq_cst seq_cst, align 4, !amdgpu.no.remote.memory [[META0]]
181 ; GFX803-NEXT: [[SUCCESS:%.*]] = extractvalue { i32, i1 } [[TMP4]], 1
182 ; GFX803-NEXT: [[NEWLOADED:%.*]] = extractvalue { i32, i1 } [[TMP4]], 0
183 ; GFX803-NEXT: [[TMP5]] = bitcast i32 [[NEWLOADED]] to <2 x half>
184 ; GFX803-NEXT: br i1 [[SUCCESS]], label [[ATOMICRMW_END:%.*]], label [[ATOMICRMW_START]]
185 ; GFX803: atomicrmw.end:
186 ; GFX803-NEXT: ret <2 x half> [[TMP5]]
187 ;
188 ; GFX906-LABEL: define <2 x half> @test_atomicrmw_fadd_v2f16_global_system__amdgpu_no_remote_memory(
189 ; GFX906-SAME: ptr addrspace(1) [[PTR:%.*]], <2 x half> [[VALUE:%.*]]) #[[ATTR0]] {
190 ; GFX906-NEXT: [[TMP1:%.*]] = load <2 x half>, ptr addrspace(1) [[PTR]], align 4
191 ; GFX906-NEXT: br label [[ATOMICRMW_START:%.*]]
192 ; GFX906: atomicrmw.start:
193 ; GFX906-NEXT: [[LOADED:%.*]] = phi <2 x half> [ [[TMP1]], [[TMP0:%.*]] ], [ [[TMP5:%.*]], [[ATOMICRMW_START]] ]
194 ; GFX906-NEXT: [[NEW:%.*]] = fadd <2 x half> [[LOADED]], [[VALUE]]
195 ; GFX906-NEXT: [[TMP2:%.*]] = bitcast <2 x half> [[NEW]] to i32
196 ; GFX906-NEXT: [[TMP3:%.*]] = bitcast <2 x half> [[LOADED]] to i32
197 ; GFX906-NEXT: [[TMP4:%.*]] = cmpxchg ptr addrspace(1) [[PTR]], i32 [[TMP3]], i32 [[TMP2]] seq_cst seq_cst, align 4, !amdgpu.no.remote.memory [[META0]]
198 ; GFX906-NEXT: [[SUCCESS:%.*]] = extractvalue { i32, i1 } [[TMP4]], 1
199 ; GFX906-NEXT: [[NEWLOADED:%.*]] = extractvalue { i32, i1 } [[TMP4]], 0
200 ; GFX906-NEXT: [[TMP5]] = bitcast i32 [[NEWLOADED]] to <2 x half>
201 ; GFX906-NEXT: br i1 [[SUCCESS]], label [[ATOMICRMW_END:%.*]], label [[ATOMICRMW_START]]
202 ; GFX906: atomicrmw.end:
203 ; GFX906-NEXT: ret <2 x half> [[TMP5]]
204 ;
205 ; GFX908-LABEL: define <2 x half> @test_atomicrmw_fadd_v2f16_global_system__amdgpu_no_remote_memory(
206 ; GFX908-SAME: ptr addrspace(1) [[PTR:%.*]], <2 x half> [[VALUE:%.*]]) #[[ATTR0]] {
207 ; GFX908-NEXT: [[TMP1:%.*]] = load <2 x half>, ptr addrspace(1) [[PTR]], align 4
208 ; GFX908-NEXT: br label [[ATOMICRMW_START:%.*]]
209 ; GFX908: atomicrmw.start:
210 ; GFX908-NEXT: [[LOADED:%.*]] = phi <2 x half> [ [[TMP1]], [[TMP0:%.*]] ], [ [[TMP5:%.*]], [[ATOMICRMW_START]] ]
211 ; GFX908-NEXT: [[NEW:%.*]] = fadd <2 x half> [[LOADED]], [[VALUE]]
212 ; GFX908-NEXT: [[TMP2:%.*]] = bitcast <2 x half> [[NEW]] to i32
213 ; GFX908-NEXT: [[TMP3:%.*]] = bitcast <2 x half> [[LOADED]] to i32
214 ; GFX908-NEXT: [[TMP4:%.*]] = cmpxchg ptr addrspace(1) [[PTR]], i32 [[TMP3]], i32 [[TMP2]] seq_cst seq_cst, align 4, !amdgpu.no.remote.memory [[META0]]
215 ; GFX908-NEXT: [[SUCCESS:%.*]] = extractvalue { i32, i1 } [[TMP4]], 1
216 ; GFX908-NEXT: [[NEWLOADED:%.*]] = extractvalue { i32, i1 } [[TMP4]], 0
217 ; GFX908-NEXT: [[TMP5]] = bitcast i32 [[NEWLOADED]] to <2 x half>
218 ; GFX908-NEXT: br i1 [[SUCCESS]], label [[ATOMICRMW_END:%.*]], label [[ATOMICRMW_START]]
219 ; GFX908: atomicrmw.end:
220 ; GFX908-NEXT: ret <2 x half> [[TMP5]]
221 ;
222 ; GFX90A-LABEL: define <2 x half> @test_atomicrmw_fadd_v2f16_global_system__amdgpu_no_remote_memory(
223 ; GFX90A-SAME: ptr addrspace(1) [[PTR:%.*]], <2 x half> [[VALUE:%.*]]) #[[ATTR0]] {
224 ; GFX90A-NEXT: [[TMP1:%.*]] = load <2 x half>, ptr addrspace(1) [[PTR]], align 4
225 ; GFX90A-NEXT: br label [[ATOMICRMW_START:%.*]]
226 ; GFX90A: atomicrmw.start:
227 ; GFX90A-NEXT: [[LOADED:%.*]] = phi <2 x half> [ [[TMP1]], [[TMP0:%.*]] ], [ [[RES:%.*]], [[ATOMICRMW_START]] ]
228 ; GFX90A-NEXT: [[NEW:%.*]] = fadd <2 x half> [[LOADED]], [[VALUE]]
229 ; GFX90A-NEXT: [[TMP2:%.*]] = bitcast <2 x half> [[NEW]] to i32
230 ; GFX90A-NEXT: [[TMP3:%.*]] = bitcast <2 x half> [[LOADED]] to i32
231 ; GFX90A-NEXT: [[TMP4:%.*]] = cmpxchg ptr addrspace(1) [[PTR]], i32 [[TMP3]], i32 [[TMP2]] seq_cst seq_cst, align 4, !amdgpu.no.remote.memory [[META0]]
232 ; GFX90A-NEXT: [[SUCCESS:%.*]] = extractvalue { i32, i1 } [[TMP4]], 1
233 ; GFX90A-NEXT: [[NEWLOADED:%.*]] = extractvalue { i32, i1 } [[TMP4]], 0
234 ; GFX90A-NEXT: [[RES]] = bitcast i32 [[NEWLOADED]] to <2 x half>
235 ; GFX90A-NEXT: br i1 [[SUCCESS]], label [[ATOMICRMW_END:%.*]], label [[ATOMICRMW_START]]
236 ; GFX90A: atomicrmw.end:
237 ; GFX90A-NEXT: ret <2 x half> [[RES]]
238 ;
239 ; GFX940-LABEL: define <2 x half> @test_atomicrmw_fadd_v2f16_global_system__amdgpu_no_remote_memory(
240 ; GFX940-SAME: ptr addrspace(1) [[PTR:%.*]], <2 x half> [[VALUE:%.*]]) #[[ATTR0]] {
241 ; GFX940-NEXT: [[RES:%.*]] = atomicrmw fadd ptr addrspace(1) [[PTR]], <2 x half> [[VALUE]] seq_cst, align 4, !amdgpu.no.remote.memory [[META0]]
242 ; GFX940-NEXT: ret <2 x half> [[RES]]
243 ;
244 ; GFX10-LABEL: define <2 x half> @test_atomicrmw_fadd_v2f16_global_system__amdgpu_no_remote_memory(
245 ; GFX10-SAME: ptr addrspace(1) [[PTR:%.*]], <2 x half> [[VALUE:%.*]]) #[[ATTR0]] {
246 ; GFX10-NEXT: [[TMP1:%.*]] = load <2 x half>, ptr addrspace(1) [[PTR]], align 4
247 ; GFX10-NEXT: br label [[ATOMICRMW_START:%.*]]
248 ; GFX10: atomicrmw.start:
249 ; GFX10-NEXT: [[LOADED:%.*]] = phi <2 x half> [ [[TMP1]], [[TMP0:%.*]] ], [ [[TMP5:%.*]], [[ATOMICRMW_START]] ]
250 ; GFX10-NEXT: [[NEW:%.*]] = fadd <2 x half> [[LOADED]], [[VALUE]]
251 ; GFX10-NEXT: [[TMP2:%.*]] = bitcast <2 x half> [[NEW]] to i32
252 ; GFX10-NEXT: [[TMP3:%.*]] = bitcast <2 x half> [[LOADED]] to i32
253 ; GFX10-NEXT: [[TMP4:%.*]] = cmpxchg ptr addrspace(1) [[PTR]], i32 [[TMP3]], i32 [[TMP2]] seq_cst seq_cst, align 4, !amdgpu.no.remote.memory [[META0]]
254 ; GFX10-NEXT: [[SUCCESS:%.*]] = extractvalue { i32, i1 } [[TMP4]], 1
255 ; GFX10-NEXT: [[NEWLOADED:%.*]] = extractvalue { i32, i1 } [[TMP4]], 0
256 ; GFX10-NEXT: [[TMP5]] = bitcast i32 [[NEWLOADED]] to <2 x half>
257 ; GFX10-NEXT: br i1 [[SUCCESS]], label [[ATOMICRMW_END:%.*]], label [[ATOMICRMW_START]]
258 ; GFX10: atomicrmw.end:
259 ; GFX10-NEXT: ret <2 x half> [[TMP5]]
260 ;
261 ; GFX11-LABEL: define <2 x half> @test_atomicrmw_fadd_v2f16_global_system__amdgpu_no_remote_memory(
262 ; GFX11-SAME: ptr addrspace(1) [[PTR:%.*]], <2 x half> [[VALUE:%.*]]) #[[ATTR0]] {
263 ; GFX11-NEXT: [[TMP1:%.*]] = load <2 x half>, ptr addrspace(1) [[PTR]], align 4
264 ; GFX11-NEXT: br label [[ATOMICRMW_START:%.*]]
265 ; GFX11: atomicrmw.start:
266 ; GFX11-NEXT: [[LOADED:%.*]] = phi <2 x half> [ [[TMP1]], [[TMP0:%.*]] ], [ [[TMP5:%.*]], [[ATOMICRMW_START]] ]
267 ; GFX11-NEXT: [[NEW:%.*]] = fadd <2 x half> [[LOADED]], [[VALUE]]
268 ; GFX11-NEXT: [[TMP2:%.*]] = bitcast <2 x half> [[NEW]] to i32
269 ; GFX11-NEXT: [[TMP3:%.*]] = bitcast <2 x half> [[LOADED]] to i32
270 ; GFX11-NEXT: [[TMP4:%.*]] = cmpxchg ptr addrspace(1) [[PTR]], i32 [[TMP3]], i32 [[TMP2]] seq_cst seq_cst, align 4, !amdgpu.no.remote.memory [[META0]]
271 ; GFX11-NEXT: [[SUCCESS:%.*]] = extractvalue { i32, i1 } [[TMP4]], 1
272 ; GFX11-NEXT: [[NEWLOADED:%.*]] = extractvalue { i32, i1 } [[TMP4]], 0
273 ; GFX11-NEXT: [[TMP5]] = bitcast i32 [[NEWLOADED]] to <2 x half>
274 ; GFX11-NEXT: br i1 [[SUCCESS]], label [[ATOMICRMW_END:%.*]], label [[ATOMICRMW_START]]
275 ; GFX11: atomicrmw.end:
276 ; GFX11-NEXT: ret <2 x half> [[TMP5]]
277 ;
278 ; GFX12-LABEL: define <2 x half> @test_atomicrmw_fadd_v2f16_global_system__amdgpu_no_remote_memory(
279 ; GFX12-SAME: ptr addrspace(1) [[PTR:%.*]], <2 x half> [[VALUE:%.*]]) #[[ATTR0]] {
280 ; GFX12-NEXT: [[RES:%.*]] = atomicrmw fadd ptr addrspace(1) [[PTR]], <2 x half> [[VALUE]] seq_cst, align 4, !amdgpu.no.remote.memory [[META0]]
281 ; GFX12-NEXT: ret <2 x half> [[RES]]
282 ;
283 %res = atomicrmw fadd ptr addrspace(1) %ptr, <2 x half> %value seq_cst, !amdgpu.no.remote.memory !0
284 ret <2 x half> %res
285 }
287 define <2 x half> @test_atomicrmw_fadd_v2f16_global_system__amdgpu_no_fine_grained_memory__amdgpu_no_remote_memory(ptr addrspace(1) %ptr, <2 x half> %value) {
288 ; GFX803-LABEL: define <2 x half> @test_atomicrmw_fadd_v2f16_global_system__amdgpu_no_fine_grained_memory__amdgpu_no_remote_memory(
289 ; GFX803-SAME: ptr addrspace(1) [[PTR:%.*]], <2 x half> [[VALUE:%.*]]) #[[ATTR0]] {
290 ; GFX803-NEXT: [[TMP1:%.*]] = load <2 x half>, ptr addrspace(1) [[PTR]], align 4
291 ; GFX803-NEXT: br label [[ATOMICRMW_START:%.*]]
292 ; GFX803: atomicrmw.start:
293 ; GFX803-NEXT: [[LOADED:%.*]] = phi <2 x half> [ [[TMP1]], [[TMP0:%.*]] ], [ [[TMP5:%.*]], [[ATOMICRMW_START]] ]
294 ; GFX803-NEXT: [[NEW:%.*]] = fadd <2 x half> [[LOADED]], [[VALUE]]
295 ; GFX803-NEXT: [[TMP2:%.*]] = bitcast <2 x half> [[NEW]] to i32
296 ; GFX803-NEXT: [[TMP3:%.*]] = bitcast <2 x half> [[LOADED]] to i32
297 ; GFX803-NEXT: [[TMP4:%.*]] = cmpxchg ptr addrspace(1) [[PTR]], i32 [[TMP3]], i32 [[TMP2]] seq_cst seq_cst, align 4, !amdgpu.no.fine.grained.memory [[META0]], !amdgpu.no.remote.memory [[META0]]
298 ; GFX803-NEXT: [[SUCCESS:%.*]] = extractvalue { i32, i1 } [[TMP4]], 1
299 ; GFX803-NEXT: [[NEWLOADED:%.*]] = extractvalue { i32, i1 } [[TMP4]], 0
300 ; GFX803-NEXT: [[TMP5]] = bitcast i32 [[NEWLOADED]] to <2 x half>
301 ; GFX803-NEXT: br i1 [[SUCCESS]], label [[ATOMICRMW_END:%.*]], label [[ATOMICRMW_START]]
302 ; GFX803: atomicrmw.end:
303 ; GFX803-NEXT: ret <2 x half> [[TMP5]]
304 ;
305 ; GFX906-LABEL: define <2 x half> @test_atomicrmw_fadd_v2f16_global_system__amdgpu_no_fine_grained_memory__amdgpu_no_remote_memory(
306 ; GFX906-SAME: ptr addrspace(1) [[PTR:%.*]], <2 x half> [[VALUE:%.*]]) #[[ATTR0]] {
307 ; GFX906-NEXT: [[TMP1:%.*]] = load <2 x half>, ptr addrspace(1) [[PTR]], align 4
308 ; GFX906-NEXT: br label [[ATOMICRMW_START:%.*]]
309 ; GFX906: atomicrmw.start:
310 ; GFX906-NEXT: [[LOADED:%.*]] = phi <2 x half> [ [[TMP1]], [[TMP0:%.*]] ], [ [[TMP5:%.*]], [[ATOMICRMW_START]] ]
311 ; GFX906-NEXT: [[NEW:%.*]] = fadd <2 x half> [[LOADED]], [[VALUE]]
312 ; GFX906-NEXT: [[TMP2:%.*]] = bitcast <2 x half> [[NEW]] to i32
313 ; GFX906-NEXT: [[TMP3:%.*]] = bitcast <2 x half> [[LOADED]] to i32
314 ; GFX906-NEXT: [[TMP4:%.*]] = cmpxchg ptr addrspace(1) [[PTR]], i32 [[TMP3]], i32 [[TMP2]] seq_cst seq_cst, align 4, !amdgpu.no.fine.grained.memory [[META0]], !amdgpu.no.remote.memory [[META0]]
315 ; GFX906-NEXT: [[SUCCESS:%.*]] = extractvalue { i32, i1 } [[TMP4]], 1
316 ; GFX906-NEXT: [[NEWLOADED:%.*]] = extractvalue { i32, i1 } [[TMP4]], 0
317 ; GFX906-NEXT: [[TMP5]] = bitcast i32 [[NEWLOADED]] to <2 x half>
318 ; GFX906-NEXT: br i1 [[SUCCESS]], label [[ATOMICRMW_END:%.*]], label [[ATOMICRMW_START]]
319 ; GFX906: atomicrmw.end:
320 ; GFX906-NEXT: ret <2 x half> [[TMP5]]
321 ;
322 ; GFX908-LABEL: define <2 x half> @test_atomicrmw_fadd_v2f16_global_system__amdgpu_no_fine_grained_memory__amdgpu_no_remote_memory(
323 ; GFX908-SAME: ptr addrspace(1) [[PTR:%.*]], <2 x half> [[VALUE:%.*]]) #[[ATTR0]] {
324 ; GFX908-NEXT: [[TMP1:%.*]] = load <2 x half>, ptr addrspace(1) [[PTR]], align 4
325 ; GFX908-NEXT: br label [[ATOMICRMW_START:%.*]]
326 ; GFX908: atomicrmw.start:
327 ; GFX908-NEXT: [[LOADED:%.*]] = phi <2 x half> [ [[TMP1]], [[TMP0:%.*]] ], [ [[TMP5:%.*]], [[ATOMICRMW_START]] ]
328 ; GFX908-NEXT: [[NEW:%.*]] = fadd <2 x half> [[LOADED]], [[VALUE]]
329 ; GFX908-NEXT: [[TMP2:%.*]] = bitcast <2 x half> [[NEW]] to i32
330 ; GFX908-NEXT: [[TMP3:%.*]] = bitcast <2 x half> [[LOADED]] to i32
331 ; GFX908-NEXT: [[TMP4:%.*]] = cmpxchg ptr addrspace(1) [[PTR]], i32 [[TMP3]], i32 [[TMP2]] seq_cst seq_cst, align 4, !amdgpu.no.fine.grained.memory [[META0]], !amdgpu.no.remote.memory [[META0]]
332 ; GFX908-NEXT: [[SUCCESS:%.*]] = extractvalue { i32, i1 } [[TMP4]], 1
333 ; GFX908-NEXT: [[NEWLOADED:%.*]] = extractvalue { i32, i1 } [[TMP4]], 0
334 ; GFX908-NEXT: [[TMP5]] = bitcast i32 [[NEWLOADED]] to <2 x half>
335 ; GFX908-NEXT: br i1 [[SUCCESS]], label [[ATOMICRMW_END:%.*]], label [[ATOMICRMW_START]]
336 ; GFX908: atomicrmw.end:
337 ; GFX908-NEXT: ret <2 x half> [[TMP5]]
338 ;
339 ; GFX90A-LABEL: define <2 x half> @test_atomicrmw_fadd_v2f16_global_system__amdgpu_no_fine_grained_memory__amdgpu_no_remote_memory(
340 ; GFX90A-SAME: ptr addrspace(1) [[PTR:%.*]], <2 x half> [[VALUE:%.*]]) #[[ATTR0]] {
341 ; GFX90A-NEXT: [[RES:%.*]] = atomicrmw fadd ptr addrspace(1) [[PTR]], <2 x half> [[VALUE]] seq_cst, align 4, !amdgpu.no.fine.grained.memory [[META0]], !amdgpu.no.remote.memory [[META0]]
342 ; GFX90A-NEXT: ret <2 x half> [[RES]]
343 ;
344 ; GFX940-LABEL: define <2 x half> @test_atomicrmw_fadd_v2f16_global_system__amdgpu_no_fine_grained_memory__amdgpu_no_remote_memory(
345 ; GFX940-SAME: ptr addrspace(1) [[PTR:%.*]], <2 x half> [[VALUE:%.*]]) #[[ATTR0]] {
346 ; GFX940-NEXT: [[RES:%.*]] = atomicrmw fadd ptr addrspace(1) [[PTR]], <2 x half> [[VALUE]] seq_cst, align 4, !amdgpu.no.fine.grained.memory [[META0]], !amdgpu.no.remote.memory [[META0]]
347 ; GFX940-NEXT: ret <2 x half> [[RES]]
348 ;
349 ; GFX10-LABEL: define <2 x half> @test_atomicrmw_fadd_v2f16_global_system__amdgpu_no_fine_grained_memory__amdgpu_no_remote_memory(
350 ; GFX10-SAME: ptr addrspace(1) [[PTR:%.*]], <2 x half> [[VALUE:%.*]]) #[[ATTR0]] {
351 ; GFX10-NEXT: [[TMP1:%.*]] = load <2 x half>, ptr addrspace(1) [[PTR]], align 4
352 ; GFX10-NEXT: br label [[ATOMICRMW_START:%.*]]
353 ; GFX10: atomicrmw.start:
354 ; GFX10-NEXT: [[LOADED:%.*]] = phi <2 x half> [ [[TMP1]], [[TMP0:%.*]] ], [ [[TMP5:%.*]], [[ATOMICRMW_START]] ]
355 ; GFX10-NEXT: [[NEW:%.*]] = fadd <2 x half> [[LOADED]], [[VALUE]]
356 ; GFX10-NEXT: [[TMP2:%.*]] = bitcast <2 x half> [[NEW]] to i32
357 ; GFX10-NEXT: [[TMP3:%.*]] = bitcast <2 x half> [[LOADED]] to i32
358 ; GFX10-NEXT: [[TMP4:%.*]] = cmpxchg ptr addrspace(1) [[PTR]], i32 [[TMP3]], i32 [[TMP2]] seq_cst seq_cst, align 4, !amdgpu.no.fine.grained.memory [[META0]], !amdgpu.no.remote.memory [[META0]]
359 ; GFX10-NEXT: [[SUCCESS:%.*]] = extractvalue { i32, i1 } [[TMP4]], 1
360 ; GFX10-NEXT: [[NEWLOADED:%.*]] = extractvalue { i32, i1 } [[TMP4]], 0
361 ; GFX10-NEXT: [[TMP5]] = bitcast i32 [[NEWLOADED]] to <2 x half>
362 ; GFX10-NEXT: br i1 [[SUCCESS]], label [[ATOMICRMW_END:%.*]], label [[ATOMICRMW_START]]
363 ; GFX10: atomicrmw.end:
364 ; GFX10-NEXT: ret <2 x half> [[TMP5]]
365 ;
366 ; GFX11-LABEL: define <2 x half> @test_atomicrmw_fadd_v2f16_global_system__amdgpu_no_fine_grained_memory__amdgpu_no_remote_memory(
367 ; GFX11-SAME: ptr addrspace(1) [[PTR:%.*]], <2 x half> [[VALUE:%.*]]) #[[ATTR0]] {
368 ; GFX11-NEXT: [[TMP1:%.*]] = load <2 x half>, ptr addrspace(1) [[PTR]], align 4
369 ; GFX11-NEXT: br label [[ATOMICRMW_START:%.*]]
370 ; GFX11: atomicrmw.start:
371 ; GFX11-NEXT: [[LOADED:%.*]] = phi <2 x half> [ [[TMP1]], [[TMP0:%.*]] ], [ [[TMP5:%.*]], [[ATOMICRMW_START]] ]
372 ; GFX11-NEXT: [[NEW:%.*]] = fadd <2 x half> [[LOADED]], [[VALUE]]
373 ; GFX11-NEXT: [[TMP2:%.*]] = bitcast <2 x half> [[NEW]] to i32
374 ; GFX11-NEXT: [[TMP3:%.*]] = bitcast <2 x half> [[LOADED]] to i32
375 ; GFX11-NEXT: [[TMP4:%.*]] = cmpxchg ptr addrspace(1) [[PTR]], i32 [[TMP3]], i32 [[TMP2]] seq_cst seq_cst, align 4, !amdgpu.no.fine.grained.memory [[META0]], !amdgpu.no.remote.memory [[META0]]
376 ; GFX11-NEXT: [[SUCCESS:%.*]] = extractvalue { i32, i1 } [[TMP4]], 1
377 ; GFX11-NEXT: [[NEWLOADED:%.*]] = extractvalue { i32, i1 } [[TMP4]], 0
378 ; GFX11-NEXT: [[TMP5]] = bitcast i32 [[NEWLOADED]] to <2 x half>
379 ; GFX11-NEXT: br i1 [[SUCCESS]], label [[ATOMICRMW_END:%.*]], label [[ATOMICRMW_START]]
380 ; GFX11: atomicrmw.end:
381 ; GFX11-NEXT: ret <2 x half> [[TMP5]]
382 ;
383 ; GFX12-LABEL: define <2 x half> @test_atomicrmw_fadd_v2f16_global_system__amdgpu_no_fine_grained_memory__amdgpu_no_remote_memory(
384 ; GFX12-SAME: ptr addrspace(1) [[PTR:%.*]], <2 x half> [[VALUE:%.*]]) #[[ATTR0]] {
385 ; GFX12-NEXT: [[RES:%.*]] = atomicrmw fadd ptr addrspace(1) [[PTR]], <2 x half> [[VALUE]] seq_cst, align 4, !amdgpu.no.fine.grained.memory [[META0]], !amdgpu.no.remote.memory [[META0]]
386 ; GFX12-NEXT: ret <2 x half> [[RES]]
387 ;
388 %res = atomicrmw fadd ptr addrspace(1) %ptr, <2 x half> %value seq_cst, !amdgpu.no.fine.grained.memory !0, !amdgpu.no.remote.memory !0
389 ret <2 x half> %res
390 }
392 define <2 x half> @test_atomicrmw_fadd_v2f16_global_system__amdgpu_no_fine_grained_memory__amdgpu_no_remote_memory___denormal_fp_mode_v2f16_daz(ptr addrspace(1) %ptr, <2 x half> %value) #0 {
393 ; GFX803-LABEL: define <2 x half> @test_atomicrmw_fadd_v2f16_global_system__amdgpu_no_fine_grained_memory__amdgpu_no_remote_memory___denormal_fp_mode_v2f16_daz(
394 ; GFX803-SAME: ptr addrspace(1) [[PTR:%.*]], <2 x half> [[VALUE:%.*]]) #[[ATTR1:[0-9]+]] {
395 ; GFX803-NEXT: [[TMP1:%.*]] = load <2 x half>, ptr addrspace(1) [[PTR]], align 4
396 ; GFX803-NEXT: br label [[ATOMICRMW_START:%.*]]
397 ; GFX803: atomicrmw.start:
398 ; GFX803-NEXT: [[LOADED:%.*]] = phi <2 x half> [ [[TMP1]], [[TMP0:%.*]] ], [ [[TMP5:%.*]], [[ATOMICRMW_START]] ]
399 ; GFX803-NEXT: [[NEW:%.*]] = fadd <2 x half> [[LOADED]], [[VALUE]]
400 ; GFX803-NEXT: [[TMP2:%.*]] = bitcast <2 x half> [[NEW]] to i32
401 ; GFX803-NEXT: [[TMP3:%.*]] = bitcast <2 x half> [[LOADED]] to i32
402 ; GFX803-NEXT: [[TMP4:%.*]] = cmpxchg ptr addrspace(1) [[PTR]], i32 [[TMP3]], i32 [[TMP2]] seq_cst seq_cst, align 4, !amdgpu.no.fine.grained.memory [[META0]], !amdgpu.no.remote.memory [[META0]]
403 ; GFX803-NEXT: [[SUCCESS:%.*]] = extractvalue { i32, i1 } [[TMP4]], 1
404 ; GFX803-NEXT: [[NEWLOADED:%.*]] = extractvalue { i32, i1 } [[TMP4]], 0
405 ; GFX803-NEXT: [[TMP5]] = bitcast i32 [[NEWLOADED]] to <2 x half>
406 ; GFX803-NEXT: br i1 [[SUCCESS]], label [[ATOMICRMW_END:%.*]], label [[ATOMICRMW_START]]
407 ; GFX803: atomicrmw.end:
408 ; GFX803-NEXT: ret <2 x half> [[TMP5]]
409 ;
410 ; GFX906-LABEL: define <2 x half> @test_atomicrmw_fadd_v2f16_global_system__amdgpu_no_fine_grained_memory__amdgpu_no_remote_memory___denormal_fp_mode_v2f16_daz(
411 ; GFX906-SAME: ptr addrspace(1) [[PTR:%.*]], <2 x half> [[VALUE:%.*]]) #[[ATTR1:[0-9]+]] {
412 ; GFX906-NEXT: [[TMP1:%.*]] = load <2 x half>, ptr addrspace(1) [[PTR]], align 4
413 ; GFX906-NEXT: br label [[ATOMICRMW_START:%.*]]
414 ; GFX906: atomicrmw.start:
415 ; GFX906-NEXT: [[LOADED:%.*]] = phi <2 x half> [ [[TMP1]], [[TMP0:%.*]] ], [ [[TMP5:%.*]], [[ATOMICRMW_START]] ]
416 ; GFX906-NEXT: [[NEW:%.*]] = fadd <2 x half> [[LOADED]], [[VALUE]]
417 ; GFX906-NEXT: [[TMP2:%.*]] = bitcast <2 x half> [[NEW]] to i32
418 ; GFX906-NEXT: [[TMP3:%.*]] = bitcast <2 x half> [[LOADED]] to i32
419 ; GFX906-NEXT: [[TMP4:%.*]] = cmpxchg ptr addrspace(1) [[PTR]], i32 [[TMP3]], i32 [[TMP2]] seq_cst seq_cst, align 4, !amdgpu.no.fine.grained.memory [[META0]], !amdgpu.no.remote.memory [[META0]]
420 ; GFX906-NEXT: [[SUCCESS:%.*]] = extractvalue { i32, i1 } [[TMP4]], 1
421 ; GFX906-NEXT: [[NEWLOADED:%.*]] = extractvalue { i32, i1 } [[TMP4]], 0
422 ; GFX906-NEXT: [[TMP5]] = bitcast i32 [[NEWLOADED]] to <2 x half>
423 ; GFX906-NEXT: br i1 [[SUCCESS]], label [[ATOMICRMW_END:%.*]], label [[ATOMICRMW_START]]
424 ; GFX906: atomicrmw.end:
425 ; GFX906-NEXT: ret <2 x half> [[TMP5]]
426 ;
427 ; GFX908-LABEL: define <2 x half> @test_atomicrmw_fadd_v2f16_global_system__amdgpu_no_fine_grained_memory__amdgpu_no_remote_memory___denormal_fp_mode_v2f16_daz(
428 ; GFX908-SAME: ptr addrspace(1) [[PTR:%.*]], <2 x half> [[VALUE:%.*]]) #[[ATTR1:[0-9]+]] {
429 ; GFX908-NEXT: [[TMP1:%.*]] = load <2 x half>, ptr addrspace(1) [[PTR]], align 4
430 ; GFX908-NEXT: br label [[ATOMICRMW_START:%.*]]
431 ; GFX908: atomicrmw.start:
432 ; GFX908-NEXT: [[LOADED:%.*]] = phi <2 x half> [ [[TMP1]], [[TMP0:%.*]] ], [ [[TMP5:%.*]], [[ATOMICRMW_START]] ]
433 ; GFX908-NEXT: [[NEW:%.*]] = fadd <2 x half> [[LOADED]], [[VALUE]]
434 ; GFX908-NEXT: [[TMP2:%.*]] = bitcast <2 x half> [[NEW]] to i32
435 ; GFX908-NEXT: [[TMP3:%.*]] = bitcast <2 x half> [[LOADED]] to i32
436 ; GFX908-NEXT: [[TMP4:%.*]] = cmpxchg ptr addrspace(1) [[PTR]], i32 [[TMP3]], i32 [[TMP2]] seq_cst seq_cst, align 4, !amdgpu.no.fine.grained.memory [[META0]], !amdgpu.no.remote.memory [[META0]]
437 ; GFX908-NEXT: [[SUCCESS:%.*]] = extractvalue { i32, i1 } [[TMP4]], 1
438 ; GFX908-NEXT: [[NEWLOADED:%.*]] = extractvalue { i32, i1 } [[TMP4]], 0
439 ; GFX908-NEXT: [[TMP5]] = bitcast i32 [[NEWLOADED]] to <2 x half>
440 ; GFX908-NEXT: br i1 [[SUCCESS]], label [[ATOMICRMW_END:%.*]], label [[ATOMICRMW_START]]
441 ; GFX908: atomicrmw.end:
442 ; GFX908-NEXT: ret <2 x half> [[TMP5]]
443 ;
444 ; GFX90A-LABEL: define <2 x half> @test_atomicrmw_fadd_v2f16_global_system__amdgpu_no_fine_grained_memory__amdgpu_no_remote_memory___denormal_fp_mode_v2f16_daz(
445 ; GFX90A-SAME: ptr addrspace(1) [[PTR:%.*]], <2 x half> [[VALUE:%.*]]) #[[ATTR1:[0-9]+]] {
446 ; GFX90A-NEXT: [[RES:%.*]] = atomicrmw fadd ptr addrspace(1) [[PTR]], <2 x half> [[VALUE]] seq_cst, align 4, !amdgpu.no.fine.grained.memory [[META0]], !amdgpu.no.remote.memory [[META0]]
447 ; GFX90A-NEXT: ret <2 x half> [[RES]]
448 ;
449 ; GFX940-LABEL: define <2 x half> @test_atomicrmw_fadd_v2f16_global_system__amdgpu_no_fine_grained_memory__amdgpu_no_remote_memory___denormal_fp_mode_v2f16_daz(
450 ; GFX940-SAME: ptr addrspace(1) [[PTR:%.*]], <2 x half> [[VALUE:%.*]]) #[[ATTR1:[0-9]+]] {
451 ; GFX940-NEXT: [[RES:%.*]] = atomicrmw fadd ptr addrspace(1) [[PTR]], <2 x half> [[VALUE]] seq_cst, align 4, !amdgpu.no.fine.grained.memory [[META0]], !amdgpu.no.remote.memory [[META0]]
452 ; GFX940-NEXT: ret <2 x half> [[RES]]
453 ;
454 ; GFX10-LABEL: define <2 x half> @test_atomicrmw_fadd_v2f16_global_system__amdgpu_no_fine_grained_memory__amdgpu_no_remote_memory___denormal_fp_mode_v2f16_daz(
455 ; GFX10-SAME: ptr addrspace(1) [[PTR:%.*]], <2 x half> [[VALUE:%.*]]) #[[ATTR1:[0-9]+]] {
456 ; GFX10-NEXT: [[TMP1:%.*]] = load <2 x half>, ptr addrspace(1) [[PTR]], align 4
457 ; GFX10-NEXT: br label [[ATOMICRMW_START:%.*]]
458 ; GFX10: atomicrmw.start:
459 ; GFX10-NEXT: [[LOADED:%.*]] = phi <2 x half> [ [[TMP1]], [[TMP0:%.*]] ], [ [[TMP5:%.*]], [[ATOMICRMW_START]] ]
460 ; GFX10-NEXT: [[NEW:%.*]] = fadd <2 x half> [[LOADED]], [[VALUE]]
461 ; GFX10-NEXT: [[TMP2:%.*]] = bitcast <2 x half> [[NEW]] to i32
462 ; GFX10-NEXT: [[TMP3:%.*]] = bitcast <2 x half> [[LOADED]] to i32
463 ; GFX10-NEXT: [[TMP4:%.*]] = cmpxchg ptr addrspace(1) [[PTR]], i32 [[TMP3]], i32 [[TMP2]] seq_cst seq_cst, align 4, !amdgpu.no.fine.grained.memory [[META0]], !amdgpu.no.remote.memory [[META0]]
464 ; GFX10-NEXT: [[SUCCESS:%.*]] = extractvalue { i32, i1 } [[TMP4]], 1
465 ; GFX10-NEXT: [[NEWLOADED:%.*]] = extractvalue { i32, i1 } [[TMP4]], 0
466 ; GFX10-NEXT: [[TMP5]] = bitcast i32 [[NEWLOADED]] to <2 x half>
467 ; GFX10-NEXT: br i1 [[SUCCESS]], label [[ATOMICRMW_END:%.*]], label [[ATOMICRMW_START]]
468 ; GFX10: atomicrmw.end:
469 ; GFX10-NEXT: ret <2 x half> [[TMP5]]
470 ;
471 ; GFX11-LABEL: define <2 x half> @test_atomicrmw_fadd_v2f16_global_system__amdgpu_no_fine_grained_memory__amdgpu_no_remote_memory___denormal_fp_mode_v2f16_daz(
472 ; GFX11-SAME: ptr addrspace(1) [[PTR:%.*]], <2 x half> [[VALUE:%.*]]) #[[ATTR1:[0-9]+]] {
473 ; GFX11-NEXT: [[TMP1:%.*]] = load <2 x half>, ptr addrspace(1) [[PTR]], align 4
474 ; GFX11-NEXT: br label [[ATOMICRMW_START:%.*]]
475 ; GFX11: atomicrmw.start:
476 ; GFX11-NEXT: [[LOADED:%.*]] = phi <2 x half> [ [[TMP1]], [[TMP0:%.*]] ], [ [[TMP5:%.*]], [[ATOMICRMW_START]] ]
477 ; GFX11-NEXT: [[NEW:%.*]] = fadd <2 x half> [[LOADED]], [[VALUE]]
478 ; GFX11-NEXT: [[TMP2:%.*]] = bitcast <2 x half> [[NEW]] to i32
479 ; GFX11-NEXT: [[TMP3:%.*]] = bitcast <2 x half> [[LOADED]] to i32
480 ; GFX11-NEXT: [[TMP4:%.*]] = cmpxchg ptr addrspace(1) [[PTR]], i32 [[TMP3]], i32 [[TMP2]] seq_cst seq_cst, align 4, !amdgpu.no.fine.grained.memory [[META0]], !amdgpu.no.remote.memory [[META0]]
481 ; GFX11-NEXT: [[SUCCESS:%.*]] = extractvalue { i32, i1 } [[TMP4]], 1
482 ; GFX11-NEXT: [[NEWLOADED:%.*]] = extractvalue { i32, i1 } [[TMP4]], 0
483 ; GFX11-NEXT: [[TMP5]] = bitcast i32 [[NEWLOADED]] to <2 x half>
484 ; GFX11-NEXT: br i1 [[SUCCESS]], label [[ATOMICRMW_END:%.*]], label [[ATOMICRMW_START]]
485 ; GFX11: atomicrmw.end:
486 ; GFX11-NEXT: ret <2 x half> [[TMP5]]
487 ;
488 ; GFX12-LABEL: define <2 x half> @test_atomicrmw_fadd_v2f16_global_system__amdgpu_no_fine_grained_memory__amdgpu_no_remote_memory___denormal_fp_mode_v2f16_daz(
489 ; GFX12-SAME: ptr addrspace(1) [[PTR:%.*]], <2 x half> [[VALUE:%.*]]) #[[ATTR1:[0-9]+]] {
490 ; GFX12-NEXT: [[RES:%.*]] = atomicrmw fadd ptr addrspace(1) [[PTR]], <2 x half> [[VALUE]] seq_cst, align 4, !amdgpu.no.fine.grained.memory [[META0]], !amdgpu.no.remote.memory [[META0]]
491 ; GFX12-NEXT: ret <2 x half> [[RES]]
492 ;
493 %res = atomicrmw fadd ptr addrspace(1) %ptr, <2 x half> %value seq_cst, !amdgpu.no.fine.grained.memory !0, !amdgpu.no.remote.memory !0
494 ret <2 x half> %res
495 }
497 define <2 x half> @test_atomicrmw_fadd_v2f16_global_system__amdgpu_no_fine_grained_memory__amdgpu_no_remote_memory___denormal_fp_mode_v2f16_dynamic(ptr addrspace(1) %ptr, <2 x half> %value) #1 {
498 ; GFX803-LABEL: define <2 x half> @test_atomicrmw_fadd_v2f16_global_system__amdgpu_no_fine_grained_memory__amdgpu_no_remote_memory___denormal_fp_mode_v2f16_dynamic(
499 ; GFX803-SAME: ptr addrspace(1) [[PTR:%.*]], <2 x half> [[VALUE:%.*]]) #[[ATTR2:[0-9]+]] {
500 ; GFX803-NEXT: [[TMP1:%.*]] = load <2 x half>, ptr addrspace(1) [[PTR]], align 4
501 ; GFX803-NEXT: br label [[ATOMICRMW_START:%.*]]
502 ; GFX803: atomicrmw.start:
503 ; GFX803-NEXT: [[LOADED:%.*]] = phi <2 x half> [ [[TMP1]], [[TMP0:%.*]] ], [ [[TMP5:%.*]], [[ATOMICRMW_START]] ]
504 ; GFX803-NEXT: [[NEW:%.*]] = fadd <2 x half> [[LOADED]], [[VALUE]]
505 ; GFX803-NEXT: [[TMP2:%.*]] = bitcast <2 x half> [[NEW]] to i32
506 ; GFX803-NEXT: [[TMP3:%.*]] = bitcast <2 x half> [[LOADED]] to i32
507 ; GFX803-NEXT: [[TMP4:%.*]] = cmpxchg ptr addrspace(1) [[PTR]], i32 [[TMP3]], i32 [[TMP2]] seq_cst seq_cst, align 4, !amdgpu.no.fine.grained.memory [[META0]], !amdgpu.no.remote.memory [[META0]]
508 ; GFX803-NEXT: [[SUCCESS:%.*]] = extractvalue { i32, i1 } [[TMP4]], 1
509 ; GFX803-NEXT: [[NEWLOADED:%.*]] = extractvalue { i32, i1 } [[TMP4]], 0
510 ; GFX803-NEXT: [[TMP5]] = bitcast i32 [[NEWLOADED]] to <2 x half>
511 ; GFX803-NEXT: br i1 [[SUCCESS]], label [[ATOMICRMW_END:%.*]], label [[ATOMICRMW_START]]
512 ; GFX803: atomicrmw.end:
513 ; GFX803-NEXT: ret <2 x half> [[TMP5]]
514 ;
515 ; GFX906-LABEL: define <2 x half> @test_atomicrmw_fadd_v2f16_global_system__amdgpu_no_fine_grained_memory__amdgpu_no_remote_memory___denormal_fp_mode_v2f16_dynamic(
516 ; GFX906-SAME: ptr addrspace(1) [[PTR:%.*]], <2 x half> [[VALUE:%.*]]) #[[ATTR2:[0-9]+]] {
517 ; GFX906-NEXT: [[TMP1:%.*]] = load <2 x half>, ptr addrspace(1) [[PTR]], align 4
518 ; GFX906-NEXT: br label [[ATOMICRMW_START:%.*]]
519 ; GFX906: atomicrmw.start:
520 ; GFX906-NEXT: [[LOADED:%.*]] = phi <2 x half> [ [[TMP1]], [[TMP0:%.*]] ], [ [[TMP5:%.*]], [[ATOMICRMW_START]] ]
521 ; GFX906-NEXT: [[NEW:%.*]] = fadd <2 x half> [[LOADED]], [[VALUE]]
522 ; GFX906-NEXT: [[TMP2:%.*]] = bitcast <2 x half> [[NEW]] to i32
523 ; GFX906-NEXT: [[TMP3:%.*]] = bitcast <2 x half> [[LOADED]] to i32
524 ; GFX906-NEXT: [[TMP4:%.*]] = cmpxchg ptr addrspace(1) [[PTR]], i32 [[TMP3]], i32 [[TMP2]] seq_cst seq_cst, align 4, !amdgpu.no.fine.grained.memory [[META0]], !amdgpu.no.remote.memory [[META0]]
525 ; GFX906-NEXT: [[SUCCESS:%.*]] = extractvalue { i32, i1 } [[TMP4]], 1
526 ; GFX906-NEXT: [[NEWLOADED:%.*]] = extractvalue { i32, i1 } [[TMP4]], 0
527 ; GFX906-NEXT: [[TMP5]] = bitcast i32 [[NEWLOADED]] to <2 x half>
528 ; GFX906-NEXT: br i1 [[SUCCESS]], label [[ATOMICRMW_END:%.*]], label [[ATOMICRMW_START]]
529 ; GFX906: atomicrmw.end:
530 ; GFX906-NEXT: ret <2 x half> [[TMP5]]
531 ;
532 ; GFX908-LABEL: define <2 x half> @test_atomicrmw_fadd_v2f16_global_system__amdgpu_no_fine_grained_memory__amdgpu_no_remote_memory___denormal_fp_mode_v2f16_dynamic(
533 ; GFX908-SAME: ptr addrspace(1) [[PTR:%.*]], <2 x half> [[VALUE:%.*]]) #[[ATTR2:[0-9]+]] {
534 ; GFX908-NEXT: [[TMP1:%.*]] = load <2 x half>, ptr addrspace(1) [[PTR]], align 4
535 ; GFX908-NEXT: br label [[ATOMICRMW_START:%.*]]
536 ; GFX908: atomicrmw.start:
537 ; GFX908-NEXT: [[LOADED:%.*]] = phi <2 x half> [ [[TMP1]], [[TMP0:%.*]] ], [ [[TMP5:%.*]], [[ATOMICRMW_START]] ]
538 ; GFX908-NEXT: [[NEW:%.*]] = fadd <2 x half> [[LOADED]], [[VALUE]]
539 ; GFX908-NEXT: [[TMP2:%.*]] = bitcast <2 x half> [[NEW]] to i32
540 ; GFX908-NEXT: [[TMP3:%.*]] = bitcast <2 x half> [[LOADED]] to i32
541 ; GFX908-NEXT: [[TMP4:%.*]] = cmpxchg ptr addrspace(1) [[PTR]], i32 [[TMP3]], i32 [[TMP2]] seq_cst seq_cst, align 4, !amdgpu.no.fine.grained.memory [[META0]], !amdgpu.no.remote.memory [[META0]]
542 ; GFX908-NEXT: [[SUCCESS:%.*]] = extractvalue { i32, i1 } [[TMP4]], 1
543 ; GFX908-NEXT: [[NEWLOADED:%.*]] = extractvalue { i32, i1 } [[TMP4]], 0
544 ; GFX908-NEXT: [[TMP5]] = bitcast i32 [[NEWLOADED]] to <2 x half>
545 ; GFX908-NEXT: br i1 [[SUCCESS]], label [[ATOMICRMW_END:%.*]], label [[ATOMICRMW_START]]
546 ; GFX908: atomicrmw.end:
547 ; GFX908-NEXT: ret <2 x half> [[TMP5]]
548 ;
549 ; GFX90A-LABEL: define <2 x half> @test_atomicrmw_fadd_v2f16_global_system__amdgpu_no_fine_grained_memory__amdgpu_no_remote_memory___denormal_fp_mode_v2f16_dynamic(
550 ; GFX90A-SAME: ptr addrspace(1) [[PTR:%.*]], <2 x half> [[VALUE:%.*]]) #[[ATTR2:[0-9]+]] {
551 ; GFX90A-NEXT: [[RES:%.*]] = atomicrmw fadd ptr addrspace(1) [[PTR]], <2 x half> [[VALUE]] seq_cst, align 4, !amdgpu.no.fine.grained.memory [[META0]], !amdgpu.no.remote.memory [[META0]]
552 ; GFX90A-NEXT: ret <2 x half> [[RES]]
553 ;
554 ; GFX940-LABEL: define <2 x half> @test_atomicrmw_fadd_v2f16_global_system__amdgpu_no_fine_grained_memory__amdgpu_no_remote_memory___denormal_fp_mode_v2f16_dynamic(
555 ; GFX940-SAME: ptr addrspace(1) [[PTR:%.*]], <2 x half> [[VALUE:%.*]]) #[[ATTR2:[0-9]+]] {
556 ; GFX940-NEXT: [[RES:%.*]] = atomicrmw fadd ptr addrspace(1) [[PTR]], <2 x half> [[VALUE]] seq_cst, align 4, !amdgpu.no.fine.grained.memory [[META0]], !amdgpu.no.remote.memory [[META0]]
557 ; GFX940-NEXT: ret <2 x half> [[RES]]
558 ;
559 ; GFX10-LABEL: define <2 x half> @test_atomicrmw_fadd_v2f16_global_system__amdgpu_no_fine_grained_memory__amdgpu_no_remote_memory___denormal_fp_mode_v2f16_dynamic(
560 ; GFX10-SAME: ptr addrspace(1) [[PTR:%.*]], <2 x half> [[VALUE:%.*]]) #[[ATTR2:[0-9]+]] {
561 ; GFX10-NEXT: [[TMP1:%.*]] = load <2 x half>, ptr addrspace(1) [[PTR]], align 4
562 ; GFX10-NEXT: br label [[ATOMICRMW_START:%.*]]
563 ; GFX10: atomicrmw.start:
564 ; GFX10-NEXT: [[LOADED:%.*]] = phi <2 x half> [ [[TMP1]], [[TMP0:%.*]] ], [ [[TMP5:%.*]], [[ATOMICRMW_START]] ]
565 ; GFX10-NEXT: [[NEW:%.*]] = fadd <2 x half> [[LOADED]], [[VALUE]]
566 ; GFX10-NEXT: [[TMP2:%.*]] = bitcast <2 x half> [[NEW]] to i32
567 ; GFX10-NEXT: [[TMP3:%.*]] = bitcast <2 x half> [[LOADED]] to i32
568 ; GFX10-NEXT: [[TMP4:%.*]] = cmpxchg ptr addrspace(1) [[PTR]], i32 [[TMP3]], i32 [[TMP2]] seq_cst seq_cst, align 4, !amdgpu.no.fine.grained.memory [[META0]], !amdgpu.no.remote.memory [[META0]]
569 ; GFX10-NEXT: [[SUCCESS:%.*]] = extractvalue { i32, i1 } [[TMP4]], 1
570 ; GFX10-NEXT: [[NEWLOADED:%.*]] = extractvalue { i32, i1 } [[TMP4]], 0
571 ; GFX10-NEXT: [[TMP5]] = bitcast i32 [[NEWLOADED]] to <2 x half>
572 ; GFX10-NEXT: br i1 [[SUCCESS]], label [[ATOMICRMW_END:%.*]], label [[ATOMICRMW_START]]
573 ; GFX10: atomicrmw.end:
574 ; GFX10-NEXT: ret <2 x half> [[TMP5]]
575 ;
576 ; GFX11-LABEL: define <2 x half> @test_atomicrmw_fadd_v2f16_global_system__amdgpu_no_fine_grained_memory__amdgpu_no_remote_memory___denormal_fp_mode_v2f16_dynamic(
577 ; GFX11-SAME: ptr addrspace(1) [[PTR:%.*]], <2 x half> [[VALUE:%.*]]) #[[ATTR2:[0-9]+]] {
578 ; GFX11-NEXT: [[TMP1:%.*]] = load <2 x half>, ptr addrspace(1) [[PTR]], align 4
579 ; GFX11-NEXT: br label [[ATOMICRMW_START:%.*]]
580 ; GFX11: atomicrmw.start:
581 ; GFX11-NEXT: [[LOADED:%.*]] = phi <2 x half> [ [[TMP1]], [[TMP0:%.*]] ], [ [[TMP5:%.*]], [[ATOMICRMW_START]] ]
582 ; GFX11-NEXT: [[NEW:%.*]] = fadd <2 x half> [[LOADED]], [[VALUE]]
583 ; GFX11-NEXT: [[TMP2:%.*]] = bitcast <2 x half> [[NEW]] to i32
584 ; GFX11-NEXT: [[TMP3:%.*]] = bitcast <2 x half> [[LOADED]] to i32
585 ; GFX11-NEXT: [[TMP4:%.*]] = cmpxchg ptr addrspace(1) [[PTR]], i32 [[TMP3]], i32 [[TMP2]] seq_cst seq_cst, align 4, !amdgpu.no.fine.grained.memory [[META0]], !amdgpu.no.remote.memory [[META0]]
586 ; GFX11-NEXT: [[SUCCESS:%.*]] = extractvalue { i32, i1 } [[TMP4]], 1
587 ; GFX11-NEXT: [[NEWLOADED:%.*]] = extractvalue { i32, i1 } [[TMP4]], 0
588 ; GFX11-NEXT: [[TMP5]] = bitcast i32 [[NEWLOADED]] to <2 x half>
589 ; GFX11-NEXT: br i1 [[SUCCESS]], label [[ATOMICRMW_END:%.*]], label [[ATOMICRMW_START]]
590 ; GFX11: atomicrmw.end:
591 ; GFX11-NEXT: ret <2 x half> [[TMP5]]
592 ;
593 ; GFX12-LABEL: define <2 x half> @test_atomicrmw_fadd_v2f16_global_system__amdgpu_no_fine_grained_memory__amdgpu_no_remote_memory___denormal_fp_mode_v2f16_dynamic(
594 ; GFX12-SAME: ptr addrspace(1) [[PTR:%.*]], <2 x half> [[VALUE:%.*]]) #[[ATTR2:[0-9]+]] {
595 ; GFX12-NEXT: [[RES:%.*]] = atomicrmw fadd ptr addrspace(1) [[PTR]], <2 x half> [[VALUE]] seq_cst, align 4, !amdgpu.no.fine.grained.memory [[META0]], !amdgpu.no.remote.memory [[META0]]
596 ; GFX12-NEXT: ret <2 x half> [[RES]]
597 ;
598 %res = atomicrmw fadd ptr addrspace(1) %ptr, <2 x half> %value seq_cst, !amdgpu.no.fine.grained.memory !0, !amdgpu.no.remote.memory !0
599 ret <2 x half> %res
600 }
602 define <2 x half> @test_atomicrmw_fadd_v2f16_global_system__amdgpu_ignore_denormal_mode(ptr addrspace(1) %ptr, <2 x half> %value) {
603 ; COMMON-LABEL: define <2 x half> @test_atomicrmw_fadd_v2f16_global_system__amdgpu_ignore_denormal_mode(
604 ; COMMON-SAME: ptr addrspace(1) [[PTR:%.*]], <2 x half> [[VALUE:%.*]]) #[[ATTR0]] {
605 ; COMMON-NEXT: [[TMP1:%.*]] = load <2 x half>, ptr addrspace(1) [[PTR]], align 4
606 ; COMMON-NEXT: br label [[ATOMICRMW_START:%.*]]
607 ; COMMON: atomicrmw.start:
608 ; COMMON-NEXT: [[LOADED:%.*]] = phi <2 x half> [ [[TMP1]], [[TMP0:%.*]] ], [ [[TMP5:%.*]], [[ATOMICRMW_START]] ]
609 ; COMMON-NEXT: [[NEW:%.*]] = fadd <2 x half> [[LOADED]], [[VALUE]]
610 ; COMMON-NEXT: [[TMP2:%.*]] = bitcast <2 x half> [[NEW]] to i32
611 ; COMMON-NEXT: [[TMP3:%.*]] = bitcast <2 x half> [[LOADED]] to i32
612 ; COMMON-NEXT: [[TMP4:%.*]] = cmpxchg ptr addrspace(1) [[PTR]], i32 [[TMP3]], i32 [[TMP2]] seq_cst seq_cst, align 4
613 ; COMMON-NEXT: [[SUCCESS:%.*]] = extractvalue { i32, i1 } [[TMP4]], 1
614 ; COMMON-NEXT: [[NEWLOADED:%.*]] = extractvalue { i32, i1 } [[TMP4]], 0
615 ; COMMON-NEXT: [[TMP5]] = bitcast i32 [[NEWLOADED]] to <2 x half>
616 ; COMMON-NEXT: br i1 [[SUCCESS]], label [[ATOMICRMW_END:%.*]], label [[ATOMICRMW_START]]
617 ; COMMON: atomicrmw.end:
618 ; COMMON-NEXT: ret <2 x half> [[TMP5]]
619 ;
620 %res = atomicrmw fadd ptr addrspace(1) %ptr, <2 x half> %value seq_cst, align 4, !amdgpu.ignore.denormal.mode !0
621 ret <2 x half> %res
622 }
624 define <2 x half> @test_atomicrmw_fadd_v2f16_global_system__amdgpu_ignore_denormal_mode__amdgpu_no_fine_grained_memory(ptr addrspace(1) %ptr, <2 x half> %value) {
625 ; GFX803-LABEL: define <2 x half> @test_atomicrmw_fadd_v2f16_global_system__amdgpu_ignore_denormal_mode__amdgpu_no_fine_grained_memory(
626 ; GFX803-SAME: ptr addrspace(1) [[PTR:%.*]], <2 x half> [[VALUE:%.*]]) #[[ATTR0]] {
627 ; GFX803-NEXT: [[TMP1:%.*]] = load <2 x half>, ptr addrspace(1) [[PTR]], align 4
628 ; GFX803-NEXT: br label [[ATOMICRMW_START:%.*]]
629 ; GFX803: atomicrmw.start:
630 ; GFX803-NEXT: [[LOADED:%.*]] = phi <2 x half> [ [[TMP1]], [[TMP0:%.*]] ], [ [[TMP5:%.*]], [[ATOMICRMW_START]] ]
631 ; GFX803-NEXT: [[NEW:%.*]] = fadd <2 x half> [[LOADED]], [[VALUE]]
632 ; GFX803-NEXT: [[TMP2:%.*]] = bitcast <2 x half> [[NEW]] to i32
633 ; GFX803-NEXT: [[TMP3:%.*]] = bitcast <2 x half> [[LOADED]] to i32
634 ; GFX803-NEXT: [[TMP4:%.*]] = cmpxchg ptr addrspace(1) [[PTR]], i32 [[TMP3]], i32 [[TMP2]] seq_cst seq_cst, align 4, !amdgpu.no.fine.grained.memory [[META0]]
635 ; GFX803-NEXT: [[SUCCESS:%.*]] = extractvalue { i32, i1 } [[TMP4]], 1
636 ; GFX803-NEXT: [[NEWLOADED:%.*]] = extractvalue { i32, i1 } [[TMP4]], 0
637 ; GFX803-NEXT: [[TMP5]] = bitcast i32 [[NEWLOADED]] to <2 x half>
638 ; GFX803-NEXT: br i1 [[SUCCESS]], label [[ATOMICRMW_END:%.*]], label [[ATOMICRMW_START]]
639 ; GFX803: atomicrmw.end:
640 ; GFX803-NEXT: ret <2 x half> [[TMP5]]
641 ;
642 ; GFX906-LABEL: define <2 x half> @test_atomicrmw_fadd_v2f16_global_system__amdgpu_ignore_denormal_mode__amdgpu_no_fine_grained_memory(
643 ; GFX906-SAME: ptr addrspace(1) [[PTR:%.*]], <2 x half> [[VALUE:%.*]]) #[[ATTR0]] {
644 ; GFX906-NEXT: [[TMP1:%.*]] = load <2 x half>, ptr addrspace(1) [[PTR]], align 4
645 ; GFX906-NEXT: br label [[ATOMICRMW_START:%.*]]
646 ; GFX906: atomicrmw.start:
647 ; GFX906-NEXT: [[LOADED:%.*]] = phi <2 x half> [ [[TMP1]], [[TMP0:%.*]] ], [ [[TMP5:%.*]], [[ATOMICRMW_START]] ]
648 ; GFX906-NEXT: [[NEW:%.*]] = fadd <2 x half> [[LOADED]], [[VALUE]]
649 ; GFX906-NEXT: [[TMP2:%.*]] = bitcast <2 x half> [[NEW]] to i32
650 ; GFX906-NEXT: [[TMP3:%.*]] = bitcast <2 x half> [[LOADED]] to i32
651 ; GFX906-NEXT: [[TMP4:%.*]] = cmpxchg ptr addrspace(1) [[PTR]], i32 [[TMP3]], i32 [[TMP2]] seq_cst seq_cst, align 4, !amdgpu.no.fine.grained.memory [[META0]]
652 ; GFX906-NEXT: [[SUCCESS:%.*]] = extractvalue { i32, i1 } [[TMP4]], 1
653 ; GFX906-NEXT: [[NEWLOADED:%.*]] = extractvalue { i32, i1 } [[TMP4]], 0
654 ; GFX906-NEXT: [[TMP5]] = bitcast i32 [[NEWLOADED]] to <2 x half>
655 ; GFX906-NEXT: br i1 [[SUCCESS]], label [[ATOMICRMW_END:%.*]], label [[ATOMICRMW_START]]
656 ; GFX906: atomicrmw.end:
657 ; GFX906-NEXT: ret <2 x half> [[TMP5]]
658 ;
659 ; GFX908-LABEL: define <2 x half> @test_atomicrmw_fadd_v2f16_global_system__amdgpu_ignore_denormal_mode__amdgpu_no_fine_grained_memory(
660 ; GFX908-SAME: ptr addrspace(1) [[PTR:%.*]], <2 x half> [[VALUE:%.*]]) #[[ATTR0]] {
661 ; GFX908-NEXT: [[TMP1:%.*]] = load <2 x half>, ptr addrspace(1) [[PTR]], align 4
662 ; GFX908-NEXT: br label [[ATOMICRMW_START:%.*]]
663 ; GFX908: atomicrmw.start:
664 ; GFX908-NEXT: [[LOADED:%.*]] = phi <2 x half> [ [[TMP1]], [[TMP0:%.*]] ], [ [[TMP5:%.*]], [[ATOMICRMW_START]] ]
665 ; GFX908-NEXT: [[NEW:%.*]] = fadd <2 x half> [[LOADED]], [[VALUE]]
666 ; GFX908-NEXT: [[TMP2:%.*]] = bitcast <2 x half> [[NEW]] to i32
667 ; GFX908-NEXT: [[TMP3:%.*]] = bitcast <2 x half> [[LOADED]] to i32
668 ; GFX908-NEXT: [[TMP4:%.*]] = cmpxchg ptr addrspace(1) [[PTR]], i32 [[TMP3]], i32 [[TMP2]] seq_cst seq_cst, align 4, !amdgpu.no.fine.grained.memory [[META0]]
669 ; GFX908-NEXT: [[SUCCESS:%.*]] = extractvalue { i32, i1 } [[TMP4]], 1
670 ; GFX908-NEXT: [[NEWLOADED:%.*]] = extractvalue { i32, i1 } [[TMP4]], 0
671 ; GFX908-NEXT: [[TMP5]] = bitcast i32 [[NEWLOADED]] to <2 x half>
672 ; GFX908-NEXT: br i1 [[SUCCESS]], label [[ATOMICRMW_END:%.*]], label [[ATOMICRMW_START]]
673 ; GFX908: atomicrmw.end:
674 ; GFX908-NEXT: ret <2 x half> [[TMP5]]
675 ;
676 ; GFX90A-LABEL: define <2 x half> @test_atomicrmw_fadd_v2f16_global_system__amdgpu_ignore_denormal_mode__amdgpu_no_fine_grained_memory(
677 ; GFX90A-SAME: ptr addrspace(1) [[PTR:%.*]], <2 x half> [[VALUE:%.*]]) #[[ATTR0]] {
678 ; GFX90A-NEXT: [[RES:%.*]] = atomicrmw fadd ptr addrspace(1) [[PTR]], <2 x half> [[VALUE]] seq_cst, align 4, !amdgpu.no.fine.grained.memory [[META0]], !amdgpu.ignore.denormal.mode [[META0]]
679 ; GFX90A-NEXT: ret <2 x half> [[RES]]
680 ;
681 ; GFX940-LABEL: define <2 x half> @test_atomicrmw_fadd_v2f16_global_system__amdgpu_ignore_denormal_mode__amdgpu_no_fine_grained_memory(
682 ; GFX940-SAME: ptr addrspace(1) [[PTR:%.*]], <2 x half> [[VALUE:%.*]]) #[[ATTR0]] {
683 ; GFX940-NEXT: [[RES:%.*]] = atomicrmw fadd ptr addrspace(1) [[PTR]], <2 x half> [[VALUE]] seq_cst, align 4, !amdgpu.no.fine.grained.memory [[META0]], !amdgpu.ignore.denormal.mode [[META0]]
684 ; GFX940-NEXT: ret <2 x half> [[RES]]
685 ;
686 ; GFX10-LABEL: define <2 x half> @test_atomicrmw_fadd_v2f16_global_system__amdgpu_ignore_denormal_mode__amdgpu_no_fine_grained_memory(
687 ; GFX10-SAME: ptr addrspace(1) [[PTR:%.*]], <2 x half> [[VALUE:%.*]]) #[[ATTR0]] {
688 ; GFX10-NEXT: [[TMP1:%.*]] = load <2 x half>, ptr addrspace(1) [[PTR]], align 4
689 ; GFX10-NEXT: br label [[ATOMICRMW_START:%.*]]
690 ; GFX10: atomicrmw.start:
691 ; GFX10-NEXT: [[LOADED:%.*]] = phi <2 x half> [ [[TMP1]], [[TMP0:%.*]] ], [ [[TMP5:%.*]], [[ATOMICRMW_START]] ]
692 ; GFX10-NEXT: [[NEW:%.*]] = fadd <2 x half> [[LOADED]], [[VALUE]]
693 ; GFX10-NEXT: [[TMP2:%.*]] = bitcast <2 x half> [[NEW]] to i32
694 ; GFX10-NEXT: [[TMP3:%.*]] = bitcast <2 x half> [[LOADED]] to i32
695 ; GFX10-NEXT: [[TMP4:%.*]] = cmpxchg ptr addrspace(1) [[PTR]], i32 [[TMP3]], i32 [[TMP2]] seq_cst seq_cst, align 4, !amdgpu.no.fine.grained.memory [[META0]]
696 ; GFX10-NEXT: [[SUCCESS:%.*]] = extractvalue { i32, i1 } [[TMP4]], 1
697 ; GFX10-NEXT: [[NEWLOADED:%.*]] = extractvalue { i32, i1 } [[TMP4]], 0
698 ; GFX10-NEXT: [[TMP5]] = bitcast i32 [[NEWLOADED]] to <2 x half>
699 ; GFX10-NEXT: br i1 [[SUCCESS]], label [[ATOMICRMW_END:%.*]], label [[ATOMICRMW_START]]
700 ; GFX10: atomicrmw.end:
701 ; GFX10-NEXT: ret <2 x half> [[TMP5]]
702 ;
703 ; GFX11-LABEL: define <2 x half> @test_atomicrmw_fadd_v2f16_global_system__amdgpu_ignore_denormal_mode__amdgpu_no_fine_grained_memory(
704 ; GFX11-SAME: ptr addrspace(1) [[PTR:%.*]], <2 x half> [[VALUE:%.*]]) #[[ATTR0]] {
705 ; GFX11-NEXT: [[TMP1:%.*]] = load <2 x half>, ptr addrspace(1) [[PTR]], align 4
706 ; GFX11-NEXT: br label [[ATOMICRMW_START:%.*]]
707 ; GFX11: atomicrmw.start:
708 ; GFX11-NEXT: [[LOADED:%.*]] = phi <2 x half> [ [[TMP1]], [[TMP0:%.*]] ], [ [[TMP5:%.*]], [[ATOMICRMW_START]] ]
709 ; GFX11-NEXT: [[NEW:%.*]] = fadd <2 x half> [[LOADED]], [[VALUE]]
710 ; GFX11-NEXT: [[TMP2:%.*]] = bitcast <2 x half> [[NEW]] to i32
711 ; GFX11-NEXT: [[TMP3:%.*]] = bitcast <2 x half> [[LOADED]] to i32
712 ; GFX11-NEXT: [[TMP4:%.*]] = cmpxchg ptr addrspace(1) [[PTR]], i32 [[TMP3]], i32 [[TMP2]] seq_cst seq_cst, align 4, !amdgpu.no.fine.grained.memory [[META0]]
713 ; GFX11-NEXT: [[SUCCESS:%.*]] = extractvalue { i32, i1 } [[TMP4]], 1
714 ; GFX11-NEXT: [[NEWLOADED:%.*]] = extractvalue { i32, i1 } [[TMP4]], 0
715 ; GFX11-NEXT: [[TMP5]] = bitcast i32 [[NEWLOADED]] to <2 x half>
716 ; GFX11-NEXT: br i1 [[SUCCESS]], label [[ATOMICRMW_END:%.*]], label [[ATOMICRMW_START]]
717 ; GFX11: atomicrmw.end:
718 ; GFX11-NEXT: ret <2 x half> [[TMP5]]
719 ;
720 ; GFX12-LABEL: define <2 x half> @test_atomicrmw_fadd_v2f16_global_system__amdgpu_ignore_denormal_mode__amdgpu_no_fine_grained_memory(
721 ; GFX12-SAME: ptr addrspace(1) [[PTR:%.*]], <2 x half> [[VALUE:%.*]]) #[[ATTR0]] {
722 ; GFX12-NEXT: [[RES:%.*]] = atomicrmw fadd ptr addrspace(1) [[PTR]], <2 x half> [[VALUE]] seq_cst, align 4, !amdgpu.no.fine.grained.memory [[META0]], !amdgpu.ignore.denormal.mode [[META0]]
723 ; GFX12-NEXT: ret <2 x half> [[RES]]
724 ;
725 %res = atomicrmw fadd ptr addrspace(1) %ptr, <2 x half> %value seq_cst, align 4, !amdgpu.no.fine.grained.memory !0, !amdgpu.ignore.denormal.mode !0
726 ret <2 x half> %res
727 }
729 define <2 x half> @test_atomicrmw_fadd_v2f16_global_system__amdgpu_ignore_denormal_mode__amdgpu_no_remote_memory(ptr addrspace(1) %ptr, <2 x half> %value) {
730 ; GFX803-LABEL: define <2 x half> @test_atomicrmw_fadd_v2f16_global_system__amdgpu_ignore_denormal_mode__amdgpu_no_remote_memory(
731 ; GFX803-SAME: ptr addrspace(1) [[PTR:%.*]], <2 x half> [[VALUE:%.*]]) #[[ATTR0]] {
732 ; GFX803-NEXT: [[TMP1:%.*]] = load <2 x half>, ptr addrspace(1) [[PTR]], align 4
733 ; GFX803-NEXT: br label [[ATOMICRMW_START:%.*]]
734 ; GFX803: atomicrmw.start:
735 ; GFX803-NEXT: [[LOADED:%.*]] = phi <2 x half> [ [[TMP1]], [[TMP0:%.*]] ], [ [[TMP5:%.*]], [[ATOMICRMW_START]] ]
736 ; GFX803-NEXT: [[NEW:%.*]] = fadd <2 x half> [[LOADED]], [[VALUE]]
737 ; GFX803-NEXT: [[TMP2:%.*]] = bitcast <2 x half> [[NEW]] to i32
738 ; GFX803-NEXT: [[TMP3:%.*]] = bitcast <2 x half> [[LOADED]] to i32
739 ; GFX803-NEXT: [[TMP4:%.*]] = cmpxchg ptr addrspace(1) [[PTR]], i32 [[TMP3]], i32 [[TMP2]] seq_cst seq_cst, align 4, !amdgpu.no.remote.memory [[META0]]
740 ; GFX803-NEXT: [[SUCCESS:%.*]] = extractvalue { i32, i1 } [[TMP4]], 1
741 ; GFX803-NEXT: [[NEWLOADED:%.*]] = extractvalue { i32, i1 } [[TMP4]], 0
742 ; GFX803-NEXT: [[TMP5]] = bitcast i32 [[NEWLOADED]] to <2 x half>
743 ; GFX803-NEXT: br i1 [[SUCCESS]], label [[ATOMICRMW_END:%.*]], label [[ATOMICRMW_START]]
744 ; GFX803: atomicrmw.end:
745 ; GFX803-NEXT: ret <2 x half> [[TMP5]]
746 ;
747 ; GFX906-LABEL: define <2 x half> @test_atomicrmw_fadd_v2f16_global_system__amdgpu_ignore_denormal_mode__amdgpu_no_remote_memory(
748 ; GFX906-SAME: ptr addrspace(1) [[PTR:%.*]], <2 x half> [[VALUE:%.*]]) #[[ATTR0]] {
749 ; GFX906-NEXT: [[TMP1:%.*]] = load <2 x half>, ptr addrspace(1) [[PTR]], align 4
750 ; GFX906-NEXT: br label [[ATOMICRMW_START:%.*]]
751 ; GFX906: atomicrmw.start:
752 ; GFX906-NEXT: [[LOADED:%.*]] = phi <2 x half> [ [[TMP1]], [[TMP0:%.*]] ], [ [[TMP5:%.*]], [[ATOMICRMW_START]] ]
753 ; GFX906-NEXT: [[NEW:%.*]] = fadd <2 x half> [[LOADED]], [[VALUE]]
754 ; GFX906-NEXT: [[TMP2:%.*]] = bitcast <2 x half> [[NEW]] to i32
755 ; GFX906-NEXT: [[TMP3:%.*]] = bitcast <2 x half> [[LOADED]] to i32
756 ; GFX906-NEXT: [[TMP4:%.*]] = cmpxchg ptr addrspace(1) [[PTR]], i32 [[TMP3]], i32 [[TMP2]] seq_cst seq_cst, align 4, !amdgpu.no.remote.memory [[META0]]
757 ; GFX906-NEXT: [[SUCCESS:%.*]] = extractvalue { i32, i1 } [[TMP4]], 1
758 ; GFX906-NEXT: [[NEWLOADED:%.*]] = extractvalue { i32, i1 } [[TMP4]], 0
759 ; GFX906-NEXT: [[TMP5]] = bitcast i32 [[NEWLOADED]] to <2 x half>
760 ; GFX906-NEXT: br i1 [[SUCCESS]], label [[ATOMICRMW_END:%.*]], label [[ATOMICRMW_START]]
761 ; GFX906: atomicrmw.end:
762 ; GFX906-NEXT: ret <2 x half> [[TMP5]]
763 ;
764 ; GFX908-LABEL: define <2 x half> @test_atomicrmw_fadd_v2f16_global_system__amdgpu_ignore_denormal_mode__amdgpu_no_remote_memory(
765 ; GFX908-SAME: ptr addrspace(1) [[PTR:%.*]], <2 x half> [[VALUE:%.*]]) #[[ATTR0]] {
766 ; GFX908-NEXT: [[TMP1:%.*]] = load <2 x half>, ptr addrspace(1) [[PTR]], align 4
767 ; GFX908-NEXT: br label [[ATOMICRMW_START:%.*]]
768 ; GFX908: atomicrmw.start:
769 ; GFX908-NEXT: [[LOADED:%.*]] = phi <2 x half> [ [[TMP1]], [[TMP0:%.*]] ], [ [[TMP5:%.*]], [[ATOMICRMW_START]] ]
770 ; GFX908-NEXT: [[NEW:%.*]] = fadd <2 x half> [[LOADED]], [[VALUE]]
771 ; GFX908-NEXT: [[TMP2:%.*]] = bitcast <2 x half> [[NEW]] to i32
772 ; GFX908-NEXT: [[TMP3:%.*]] = bitcast <2 x half> [[LOADED]] to i32
773 ; GFX908-NEXT: [[TMP4:%.*]] = cmpxchg ptr addrspace(1) [[PTR]], i32 [[TMP3]], i32 [[TMP2]] seq_cst seq_cst, align 4, !amdgpu.no.remote.memory [[META0]]
774 ; GFX908-NEXT: [[SUCCESS:%.*]] = extractvalue { i32, i1 } [[TMP4]], 1
775 ; GFX908-NEXT: [[NEWLOADED:%.*]] = extractvalue { i32, i1 } [[TMP4]], 0
776 ; GFX908-NEXT: [[TMP5]] = bitcast i32 [[NEWLOADED]] to <2 x half>
777 ; GFX908-NEXT: br i1 [[SUCCESS]], label [[ATOMICRMW_END:%.*]], label [[ATOMICRMW_START]]
778 ; GFX908: atomicrmw.end:
779 ; GFX908-NEXT: ret <2 x half> [[TMP5]]
780 ;
781 ; GFX90A-LABEL: define <2 x half> @test_atomicrmw_fadd_v2f16_global_system__amdgpu_ignore_denormal_mode__amdgpu_no_remote_memory(
782 ; GFX90A-SAME: ptr addrspace(1) [[PTR:%.*]], <2 x half> [[VALUE:%.*]]) #[[ATTR0]] {
783 ; GFX90A-NEXT: [[TMP1:%.*]] = load <2 x half>, ptr addrspace(1) [[PTR]], align 4
784 ; GFX90A-NEXT: br label [[ATOMICRMW_START:%.*]]
785 ; GFX90A: atomicrmw.start:
786 ; GFX90A-NEXT: [[LOADED:%.*]] = phi <2 x half> [ [[TMP1]], [[TMP0:%.*]] ], [ [[RES:%.*]], [[ATOMICRMW_START]] ]
787 ; GFX90A-NEXT: [[NEW:%.*]] = fadd <2 x half> [[LOADED]], [[VALUE]]
788 ; GFX90A-NEXT: [[TMP2:%.*]] = bitcast <2 x half> [[NEW]] to i32
789 ; GFX90A-NEXT: [[TMP3:%.*]] = bitcast <2 x half> [[LOADED]] to i32
790 ; GFX90A-NEXT: [[TMP4:%.*]] = cmpxchg ptr addrspace(1) [[PTR]], i32 [[TMP3]], i32 [[TMP2]] seq_cst seq_cst, align 4, !amdgpu.no.remote.memory [[META0]]
791 ; GFX90A-NEXT: [[SUCCESS:%.*]] = extractvalue { i32, i1 } [[TMP4]], 1
792 ; GFX90A-NEXT: [[NEWLOADED:%.*]] = extractvalue { i32, i1 } [[TMP4]], 0
793 ; GFX90A-NEXT: [[RES]] = bitcast i32 [[NEWLOADED]] to <2 x half>
794 ; GFX90A-NEXT: br i1 [[SUCCESS]], label [[ATOMICRMW_END:%.*]], label [[ATOMICRMW_START]]
795 ; GFX90A: atomicrmw.end:
796 ; GFX90A-NEXT: ret <2 x half> [[RES]]
797 ;
798 ; GFX940-LABEL: define <2 x half> @test_atomicrmw_fadd_v2f16_global_system__amdgpu_ignore_denormal_mode__amdgpu_no_remote_memory(
799 ; GFX940-SAME: ptr addrspace(1) [[PTR:%.*]], <2 x half> [[VALUE:%.*]]) #[[ATTR0]] {
800 ; GFX940-NEXT: [[RES:%.*]] = atomicrmw fadd ptr addrspace(1) [[PTR]], <2 x half> [[VALUE]] seq_cst, align 4, !amdgpu.no.remote.memory [[META0]], !amdgpu.ignore.denormal.mode [[META0]]
801 ; GFX940-NEXT: ret <2 x half> [[RES]]
802 ;
803 ; GFX10-LABEL: define <2 x half> @test_atomicrmw_fadd_v2f16_global_system__amdgpu_ignore_denormal_mode__amdgpu_no_remote_memory(
804 ; GFX10-SAME: ptr addrspace(1) [[PTR:%.*]], <2 x half> [[VALUE:%.*]]) #[[ATTR0]] {
805 ; GFX10-NEXT: [[TMP1:%.*]] = load <2 x half>, ptr addrspace(1) [[PTR]], align 4
806 ; GFX10-NEXT: br label [[ATOMICRMW_START:%.*]]
807 ; GFX10: atomicrmw.start:
808 ; GFX10-NEXT: [[LOADED:%.*]] = phi <2 x half> [ [[TMP1]], [[TMP0:%.*]] ], [ [[TMP5:%.*]], [[ATOMICRMW_START]] ]
809 ; GFX10-NEXT: [[NEW:%.*]] = fadd <2 x half> [[LOADED]], [[VALUE]]
810 ; GFX10-NEXT: [[TMP2:%.*]] = bitcast <2 x half> [[NEW]] to i32
811 ; GFX10-NEXT: [[TMP3:%.*]] = bitcast <2 x half> [[LOADED]] to i32
812 ; GFX10-NEXT: [[TMP4:%.*]] = cmpxchg ptr addrspace(1) [[PTR]], i32 [[TMP3]], i32 [[TMP2]] seq_cst seq_cst, align 4, !amdgpu.no.remote.memory [[META0]]
813 ; GFX10-NEXT: [[SUCCESS:%.*]] = extractvalue { i32, i1 } [[TMP4]], 1
814 ; GFX10-NEXT: [[NEWLOADED:%.*]] = extractvalue { i32, i1 } [[TMP4]], 0
815 ; GFX10-NEXT: [[TMP5]] = bitcast i32 [[NEWLOADED]] to <2 x half>
816 ; GFX10-NEXT: br i1 [[SUCCESS]], label [[ATOMICRMW_END:%.*]], label [[ATOMICRMW_START]]
817 ; GFX10: atomicrmw.end:
818 ; GFX10-NEXT: ret <2 x half> [[TMP5]]
819 ;
820 ; GFX11-LABEL: define <2 x half> @test_atomicrmw_fadd_v2f16_global_system__amdgpu_ignore_denormal_mode__amdgpu_no_remote_memory(
821 ; GFX11-SAME: ptr addrspace(1) [[PTR:%.*]], <2 x half> [[VALUE:%.*]]) #[[ATTR0]] {
822 ; GFX11-NEXT: [[TMP1:%.*]] = load <2 x half>, ptr addrspace(1) [[PTR]], align 4
823 ; GFX11-NEXT: br label [[ATOMICRMW_START:%.*]]
824 ; GFX11: atomicrmw.start:
825 ; GFX11-NEXT: [[LOADED:%.*]] = phi <2 x half> [ [[TMP1]], [[TMP0:%.*]] ], [ [[TMP5:%.*]], [[ATOMICRMW_START]] ]
826 ; GFX11-NEXT: [[NEW:%.*]] = fadd <2 x half> [[LOADED]], [[VALUE]]
827 ; GFX11-NEXT: [[TMP2:%.*]] = bitcast <2 x half> [[NEW]] to i32
828 ; GFX11-NEXT: [[TMP3:%.*]] = bitcast <2 x half> [[LOADED]] to i32
829 ; GFX11-NEXT: [[TMP4:%.*]] = cmpxchg ptr addrspace(1) [[PTR]], i32 [[TMP3]], i32 [[TMP2]] seq_cst seq_cst, align 4, !amdgpu.no.remote.memory [[META0]]
830 ; GFX11-NEXT: [[SUCCESS:%.*]] = extractvalue { i32, i1 } [[TMP4]], 1
831 ; GFX11-NEXT: [[NEWLOADED:%.*]] = extractvalue { i32, i1 } [[TMP4]], 0
832 ; GFX11-NEXT: [[TMP5]] = bitcast i32 [[NEWLOADED]] to <2 x half>
833 ; GFX11-NEXT: br i1 [[SUCCESS]], label [[ATOMICRMW_END:%.*]], label [[ATOMICRMW_START]]
834 ; GFX11: atomicrmw.end:
835 ; GFX11-NEXT: ret <2 x half> [[TMP5]]
836 ;
837 ; GFX12-LABEL: define <2 x half> @test_atomicrmw_fadd_v2f16_global_system__amdgpu_ignore_denormal_mode__amdgpu_no_remote_memory(
838 ; GFX12-SAME: ptr addrspace(1) [[PTR:%.*]], <2 x half> [[VALUE:%.*]]) #[[ATTR0]] {
839 ; GFX12-NEXT: [[RES:%.*]] = atomicrmw fadd ptr addrspace(1) [[PTR]], <2 x half> [[VALUE]] seq_cst, align 4, !amdgpu.no.remote.memory [[META0]], !amdgpu.ignore.denormal.mode [[META0]]
840 ; GFX12-NEXT: ret <2 x half> [[RES]]
841 ;
842 %res = atomicrmw fadd ptr addrspace(1) %ptr, <2 x half> %value seq_cst, align 4, !amdgpu.no.remote.memory !0, !amdgpu.ignore.denormal.mode !0
843 ret <2 x half> %res
844 }
846 define <2 x half> @test_atomicrmw_fadd_v2f16_global_system__amdgpu_ignore_denormal_mode__amdgpu_no_fine_grained_memory__amdgpu_no_remote_memory(ptr addrspace(1) %ptr, <2 x half> %value) {
847 ; GFX803-LABEL: define <2 x half> @test_atomicrmw_fadd_v2f16_global_system__amdgpu_ignore_denormal_mode__amdgpu_no_fine_grained_memory__amdgpu_no_remote_memory(
848 ; GFX803-SAME: ptr addrspace(1) [[PTR:%.*]], <2 x half> [[VALUE:%.*]]) #[[ATTR0]] {
849 ; GFX803-NEXT: [[TMP1:%.*]] = load <2 x half>, ptr addrspace(1) [[PTR]], align 4
850 ; GFX803-NEXT: br label [[ATOMICRMW_START:%.*]]
851 ; GFX803: atomicrmw.start:
852 ; GFX803-NEXT: [[LOADED:%.*]] = phi <2 x half> [ [[TMP1]], [[TMP0:%.*]] ], [ [[TMP5:%.*]], [[ATOMICRMW_START]] ]
853 ; GFX803-NEXT: [[NEW:%.*]] = fadd <2 x half> [[LOADED]], [[VALUE]]
854 ; GFX803-NEXT: [[TMP2:%.*]] = bitcast <2 x half> [[NEW]] to i32
855 ; GFX803-NEXT: [[TMP3:%.*]] = bitcast <2 x half> [[LOADED]] to i32
856 ; GFX803-NEXT: [[TMP4:%.*]] = cmpxchg ptr addrspace(1) [[PTR]], i32 [[TMP3]], i32 [[TMP2]] seq_cst seq_cst, align 4, !amdgpu.no.fine.grained.memory [[META0]], !amdgpu.no.remote.memory [[META0]]
857 ; GFX803-NEXT: [[SUCCESS:%.*]] = extractvalue { i32, i1 } [[TMP4]], 1
858 ; GFX803-NEXT: [[NEWLOADED:%.*]] = extractvalue { i32, i1 } [[TMP4]], 0
859 ; GFX803-NEXT: [[TMP5]] = bitcast i32 [[NEWLOADED]] to <2 x half>
860 ; GFX803-NEXT: br i1 [[SUCCESS]], label [[ATOMICRMW_END:%.*]], label [[ATOMICRMW_START]]
861 ; GFX803: atomicrmw.end:
862 ; GFX803-NEXT: ret <2 x half> [[TMP5]]
863 ;
864 ; GFX906-LABEL: define <2 x half> @test_atomicrmw_fadd_v2f16_global_system__amdgpu_ignore_denormal_mode__amdgpu_no_fine_grained_memory__amdgpu_no_remote_memory(
865 ; GFX906-SAME: ptr addrspace(1) [[PTR:%.*]], <2 x half> [[VALUE:%.*]]) #[[ATTR0]] {
866 ; GFX906-NEXT: [[TMP1:%.*]] = load <2 x half>, ptr addrspace(1) [[PTR]], align 4
867 ; GFX906-NEXT: br label [[ATOMICRMW_START:%.*]]
868 ; GFX906: atomicrmw.start:
869 ; GFX906-NEXT: [[LOADED:%.*]] = phi <2 x half> [ [[TMP1]], [[TMP0:%.*]] ], [ [[TMP5:%.*]], [[ATOMICRMW_START]] ]
870 ; GFX906-NEXT: [[NEW:%.*]] = fadd <2 x half> [[LOADED]], [[VALUE]]
871 ; GFX906-NEXT: [[TMP2:%.*]] = bitcast <2 x half> [[NEW]] to i32
872 ; GFX906-NEXT: [[TMP3:%.*]] = bitcast <2 x half> [[LOADED]] to i32
873 ; GFX906-NEXT: [[TMP4:%.*]] = cmpxchg ptr addrspace(1) [[PTR]], i32 [[TMP3]], i32 [[TMP2]] seq_cst seq_cst, align 4, !amdgpu.no.fine.grained.memory [[META0]], !amdgpu.no.remote.memory [[META0]]
874 ; GFX906-NEXT: [[SUCCESS:%.*]] = extractvalue { i32, i1 } [[TMP4]], 1
875 ; GFX906-NEXT: [[NEWLOADED:%.*]] = extractvalue { i32, i1 } [[TMP4]], 0
876 ; GFX906-NEXT: [[TMP5]] = bitcast i32 [[NEWLOADED]] to <2 x half>
877 ; GFX906-NEXT: br i1 [[SUCCESS]], label [[ATOMICRMW_END:%.*]], label [[ATOMICRMW_START]]
878 ; GFX906: atomicrmw.end:
879 ; GFX906-NEXT: ret <2 x half> [[TMP5]]
880 ;
881 ; GFX908-LABEL: define <2 x half> @test_atomicrmw_fadd_v2f16_global_system__amdgpu_ignore_denormal_mode__amdgpu_no_fine_grained_memory__amdgpu_no_remote_memory(
882 ; GFX908-SAME: ptr addrspace(1) [[PTR:%.*]], <2 x half> [[VALUE:%.*]]) #[[ATTR0]] {
883 ; GFX908-NEXT: [[TMP1:%.*]] = load <2 x half>, ptr addrspace(1) [[PTR]], align 4
884 ; GFX908-NEXT: br label [[ATOMICRMW_START:%.*]]
885 ; GFX908: atomicrmw.start:
886 ; GFX908-NEXT: [[LOADED:%.*]] = phi <2 x half> [ [[TMP1]], [[TMP0:%.*]] ], [ [[TMP5:%.*]], [[ATOMICRMW_START]] ]
887 ; GFX908-NEXT: [[NEW:%.*]] = fadd <2 x half> [[LOADED]], [[VALUE]]
888 ; GFX908-NEXT: [[TMP2:%.*]] = bitcast <2 x half> [[NEW]] to i32
889 ; GFX908-NEXT: [[TMP3:%.*]] = bitcast <2 x half> [[LOADED]] to i32
890 ; GFX908-NEXT: [[TMP4:%.*]] = cmpxchg ptr addrspace(1) [[PTR]], i32 [[TMP3]], i32 [[TMP2]] seq_cst seq_cst, align 4, !amdgpu.no.fine.grained.memory [[META0]], !amdgpu.no.remote.memory [[META0]]
891 ; GFX908-NEXT: [[SUCCESS:%.*]] = extractvalue { i32, i1 } [[TMP4]], 1
892 ; GFX908-NEXT: [[NEWLOADED:%.*]] = extractvalue { i32, i1 } [[TMP4]], 0
893 ; GFX908-NEXT: [[TMP5]] = bitcast i32 [[NEWLOADED]] to <2 x half>
894 ; GFX908-NEXT: br i1 [[SUCCESS]], label [[ATOMICRMW_END:%.*]], label [[ATOMICRMW_START]]
895 ; GFX908: atomicrmw.end:
896 ; GFX908-NEXT: ret <2 x half> [[TMP5]]
897 ;
898 ; GFX90A-LABEL: define <2 x half> @test_atomicrmw_fadd_v2f16_global_system__amdgpu_ignore_denormal_mode__amdgpu_no_fine_grained_memory__amdgpu_no_remote_memory(
899 ; GFX90A-SAME: ptr addrspace(1) [[PTR:%.*]], <2 x half> [[VALUE:%.*]]) #[[ATTR0]] {
900 ; GFX90A-NEXT: [[RES:%.*]] = atomicrmw fadd ptr addrspace(1) [[PTR]], <2 x half> [[VALUE]] seq_cst, align 4, !amdgpu.no.fine.grained.memory [[META0]], !amdgpu.no.remote.memory [[META0]], !amdgpu.ignore.denormal.mode [[META0]]
901 ; GFX90A-NEXT: ret <2 x half> [[RES]]
902 ;
903 ; GFX940-LABEL: define <2 x half> @test_atomicrmw_fadd_v2f16_global_system__amdgpu_ignore_denormal_mode__amdgpu_no_fine_grained_memory__amdgpu_no_remote_memory(
904 ; GFX940-SAME: ptr addrspace(1) [[PTR:%.*]], <2 x half> [[VALUE:%.*]]) #[[ATTR0]] {
905 ; GFX940-NEXT: [[RES:%.*]] = atomicrmw fadd ptr addrspace(1) [[PTR]], <2 x half> [[VALUE]] seq_cst, align 4, !amdgpu.no.fine.grained.memory [[META0]], !amdgpu.no.remote.memory [[META0]], !amdgpu.ignore.denormal.mode [[META0]]
906 ; GFX940-NEXT: ret <2 x half> [[RES]]
907 ;
908 ; GFX10-LABEL: define <2 x half> @test_atomicrmw_fadd_v2f16_global_system__amdgpu_ignore_denormal_mode__amdgpu_no_fine_grained_memory__amdgpu_no_remote_memory(
909 ; GFX10-SAME: ptr addrspace(1) [[PTR:%.*]], <2 x half> [[VALUE:%.*]]) #[[ATTR0]] {
910 ; GFX10-NEXT: [[TMP1:%.*]] = load <2 x half>, ptr addrspace(1) [[PTR]], align 4
911 ; GFX10-NEXT: br label [[ATOMICRMW_START:%.*]]
912 ; GFX10: atomicrmw.start:
913 ; GFX10-NEXT: [[LOADED:%.*]] = phi <2 x half> [ [[TMP1]], [[TMP0:%.*]] ], [ [[TMP5:%.*]], [[ATOMICRMW_START]] ]
914 ; GFX10-NEXT: [[NEW:%.*]] = fadd <2 x half> [[LOADED]], [[VALUE]]
915 ; GFX10-NEXT: [[TMP2:%.*]] = bitcast <2 x half> [[NEW]] to i32
916 ; GFX10-NEXT: [[TMP3:%.*]] = bitcast <2 x half> [[LOADED]] to i32
917 ; GFX10-NEXT: [[TMP4:%.*]] = cmpxchg ptr addrspace(1) [[PTR]], i32 [[TMP3]], i32 [[TMP2]] seq_cst seq_cst, align 4, !amdgpu.no.fine.grained.memory [[META0]], !amdgpu.no.remote.memory [[META0]]
918 ; GFX10-NEXT: [[SUCCESS:%.*]] = extractvalue { i32, i1 } [[TMP4]], 1
919 ; GFX10-NEXT: [[NEWLOADED:%.*]] = extractvalue { i32, i1 } [[TMP4]], 0
920 ; GFX10-NEXT: [[TMP5]] = bitcast i32 [[NEWLOADED]] to <2 x half>
921 ; GFX10-NEXT: br i1 [[SUCCESS]], label [[ATOMICRMW_END:%.*]], label [[ATOMICRMW_START]]
922 ; GFX10: atomicrmw.end:
923 ; GFX10-NEXT: ret <2 x half> [[TMP5]]
924 ;
925 ; GFX11-LABEL: define <2 x half> @test_atomicrmw_fadd_v2f16_global_system__amdgpu_ignore_denormal_mode__amdgpu_no_fine_grained_memory__amdgpu_no_remote_memory(
926 ; GFX11-SAME: ptr addrspace(1) [[PTR:%.*]], <2 x half> [[VALUE:%.*]]) #[[ATTR0]] {
927 ; GFX11-NEXT: [[TMP1:%.*]] = load <2 x half>, ptr addrspace(1) [[PTR]], align 4
928 ; GFX11-NEXT: br label [[ATOMICRMW_START:%.*]]
929 ; GFX11: atomicrmw.start:
930 ; GFX11-NEXT: [[LOADED:%.*]] = phi <2 x half> [ [[TMP1]], [[TMP0:%.*]] ], [ [[TMP5:%.*]], [[ATOMICRMW_START]] ]
931 ; GFX11-NEXT: [[NEW:%.*]] = fadd <2 x half> [[LOADED]], [[VALUE]]
932 ; GFX11-NEXT: [[TMP2:%.*]] = bitcast <2 x half> [[NEW]] to i32
933 ; GFX11-NEXT: [[TMP3:%.*]] = bitcast <2 x half> [[LOADED]] to i32
934 ; GFX11-NEXT: [[TMP4:%.*]] = cmpxchg ptr addrspace(1) [[PTR]], i32 [[TMP3]], i32 [[TMP2]] seq_cst seq_cst, align 4, !amdgpu.no.fine.grained.memory [[META0]], !amdgpu.no.remote.memory [[META0]]
935 ; GFX11-NEXT: [[SUCCESS:%.*]] = extractvalue { i32, i1 } [[TMP4]], 1
936 ; GFX11-NEXT: [[NEWLOADED:%.*]] = extractvalue { i32, i1 } [[TMP4]], 0
937 ; GFX11-NEXT: [[TMP5]] = bitcast i32 [[NEWLOADED]] to <2 x half>
938 ; GFX11-NEXT: br i1 [[SUCCESS]], label [[ATOMICRMW_END:%.*]], label [[ATOMICRMW_START]]
939 ; GFX11: atomicrmw.end:
940 ; GFX11-NEXT: ret <2 x half> [[TMP5]]
941 ;
942 ; GFX12-LABEL: define <2 x half> @test_atomicrmw_fadd_v2f16_global_system__amdgpu_ignore_denormal_mode__amdgpu_no_fine_grained_memory__amdgpu_no_remote_memory(
943 ; GFX12-SAME: ptr addrspace(1) [[PTR:%.*]], <2 x half> [[VALUE:%.*]]) #[[ATTR0]] {
944 ; GFX12-NEXT: [[RES:%.*]] = atomicrmw fadd ptr addrspace(1) [[PTR]], <2 x half> [[VALUE]] seq_cst, align 4, !amdgpu.no.fine.grained.memory [[META0]], !amdgpu.no.remote.memory [[META0]], !amdgpu.ignore.denormal.mode [[META0]]
945 ; GFX12-NEXT: ret <2 x half> [[RES]]
946 ;
947 %res = atomicrmw fadd ptr addrspace(1) %ptr, <2 x half> %value seq_cst, align 4, !amdgpu.no.fine.grained.memory !0, !amdgpu.no.remote.memory !0, !amdgpu.ignore.denormal.mode !0
948 ret <2 x half> %res
949 }
951 define <2 x half> @test_atomicrmw_fadd_v2f16_global_system__amdgpu_ignore_denormal_mode__amdgpu_no_fine_grained_memory__amdgpu_no_remote_memory__denormal_mode_daz(ptr addrspace(1) %ptr, <2 x half> %value) #0 {
952 ; GFX803-LABEL: define <2 x half> @test_atomicrmw_fadd_v2f16_global_system__amdgpu_ignore_denormal_mode__amdgpu_no_fine_grained_memory__amdgpu_no_remote_memory__denormal_mode_daz(
953 ; GFX803-SAME: ptr addrspace(1) [[PTR:%.*]], <2 x half> [[VALUE:%.*]]) #[[ATTR1]] {
954 ; GFX803-NEXT: [[TMP1:%.*]] = load <2 x half>, ptr addrspace(1) [[PTR]], align 4
955 ; GFX803-NEXT: br label [[ATOMICRMW_START:%.*]]
956 ; GFX803: atomicrmw.start:
957 ; GFX803-NEXT: [[LOADED:%.*]] = phi <2 x half> [ [[TMP1]], [[TMP0:%.*]] ], [ [[TMP5:%.*]], [[ATOMICRMW_START]] ]
958 ; GFX803-NEXT: [[NEW:%.*]] = fadd <2 x half> [[LOADED]], [[VALUE]]
959 ; GFX803-NEXT: [[TMP2:%.*]] = bitcast <2 x half> [[NEW]] to i32
960 ; GFX803-NEXT: [[TMP3:%.*]] = bitcast <2 x half> [[LOADED]] to i32
961 ; GFX803-NEXT: [[TMP4:%.*]] = cmpxchg ptr addrspace(1) [[PTR]], i32 [[TMP3]], i32 [[TMP2]] seq_cst seq_cst, align 4, !amdgpu.no.fine.grained.memory [[META0]], !amdgpu.no.remote.memory [[META0]]
962 ; GFX803-NEXT: [[SUCCESS:%.*]] = extractvalue { i32, i1 } [[TMP4]], 1
963 ; GFX803-NEXT: [[NEWLOADED:%.*]] = extractvalue { i32, i1 } [[TMP4]], 0
964 ; GFX803-NEXT: [[TMP5]] = bitcast i32 [[NEWLOADED]] to <2 x half>
965 ; GFX803-NEXT: br i1 [[SUCCESS]], label [[ATOMICRMW_END:%.*]], label [[ATOMICRMW_START]]
966 ; GFX803: atomicrmw.end:
967 ; GFX803-NEXT: ret <2 x half> [[TMP5]]
968 ;
969 ; GFX906-LABEL: define <2 x half> @test_atomicrmw_fadd_v2f16_global_system__amdgpu_ignore_denormal_mode__amdgpu_no_fine_grained_memory__amdgpu_no_remote_memory__denormal_mode_daz(
970 ; GFX906-SAME: ptr addrspace(1) [[PTR:%.*]], <2 x half> [[VALUE:%.*]]) #[[ATTR1]] {
971 ; GFX906-NEXT: [[TMP1:%.*]] = load <2 x half>, ptr addrspace(1) [[PTR]], align 4
972 ; GFX906-NEXT: br label [[ATOMICRMW_START:%.*]]
973 ; GFX906: atomicrmw.start:
974 ; GFX906-NEXT: [[LOADED:%.*]] = phi <2 x half> [ [[TMP1]], [[TMP0:%.*]] ], [ [[TMP5:%.*]], [[ATOMICRMW_START]] ]
975 ; GFX906-NEXT: [[NEW:%.*]] = fadd <2 x half> [[LOADED]], [[VALUE]]
976 ; GFX906-NEXT: [[TMP2:%.*]] = bitcast <2 x half> [[NEW]] to i32
977 ; GFX906-NEXT: [[TMP3:%.*]] = bitcast <2 x half> [[LOADED]] to i32
978 ; GFX906-NEXT: [[TMP4:%.*]] = cmpxchg ptr addrspace(1) [[PTR]], i32 [[TMP3]], i32 [[TMP2]] seq_cst seq_cst, align 4, !amdgpu.no.fine.grained.memory [[META0]], !amdgpu.no.remote.memory [[META0]]
979 ; GFX906-NEXT: [[SUCCESS:%.*]] = extractvalue { i32, i1 } [[TMP4]], 1
980 ; GFX906-NEXT: [[NEWLOADED:%.*]] = extractvalue { i32, i1 } [[TMP4]], 0
981 ; GFX906-NEXT: [[TMP5]] = bitcast i32 [[NEWLOADED]] to <2 x half>
982 ; GFX906-NEXT: br i1 [[SUCCESS]], label [[ATOMICRMW_END:%.*]], label [[ATOMICRMW_START]]
983 ; GFX906: atomicrmw.end:
984 ; GFX906-NEXT: ret <2 x half> [[TMP5]]
985 ;
986 ; GFX908-LABEL: define <2 x half> @test_atomicrmw_fadd_v2f16_global_system__amdgpu_ignore_denormal_mode__amdgpu_no_fine_grained_memory__amdgpu_no_remote_memory__denormal_mode_daz(
987 ; GFX908-SAME: ptr addrspace(1) [[PTR:%.*]], <2 x half> [[VALUE:%.*]]) #[[ATTR1]] {
988 ; GFX908-NEXT: [[TMP1:%.*]] = load <2 x half>, ptr addrspace(1) [[PTR]], align 4
989 ; GFX908-NEXT: br label [[ATOMICRMW_START:%.*]]
990 ; GFX908: atomicrmw.start:
991 ; GFX908-NEXT: [[LOADED:%.*]] = phi <2 x half> [ [[TMP1]], [[TMP0:%.*]] ], [ [[TMP5:%.*]], [[ATOMICRMW_START]] ]
992 ; GFX908-NEXT: [[NEW:%.*]] = fadd <2 x half> [[LOADED]], [[VALUE]]
993 ; GFX908-NEXT: [[TMP2:%.*]] = bitcast <2 x half> [[NEW]] to i32
994 ; GFX908-NEXT: [[TMP3:%.*]] = bitcast <2 x half> [[LOADED]] to i32
995 ; GFX908-NEXT: [[TMP4:%.*]] = cmpxchg ptr addrspace(1) [[PTR]], i32 [[TMP3]], i32 [[TMP2]] seq_cst seq_cst, align 4, !amdgpu.no.fine.grained.memory [[META0]], !amdgpu.no.remote.memory [[META0]]
996 ; GFX908-NEXT: [[SUCCESS:%.*]] = extractvalue { i32, i1 } [[TMP4]], 1
997 ; GFX908-NEXT: [[NEWLOADED:%.*]] = extractvalue { i32, i1 } [[TMP4]], 0
998 ; GFX908-NEXT: [[TMP5]] = bitcast i32 [[NEWLOADED]] to <2 x half>
999 ; GFX908-NEXT: br i1 [[SUCCESS]], label [[ATOMICRMW_END:%.*]], label [[ATOMICRMW_START]]
1000 ; GFX908: atomicrmw.end:
1001 ; GFX908-NEXT: ret <2 x half> [[TMP5]]
1002 ;
1003 ; GFX90A-LABEL: define <2 x half> @test_atomicrmw_fadd_v2f16_global_system__amdgpu_ignore_denormal_mode__amdgpu_no_fine_grained_memory__amdgpu_no_remote_memory__denormal_mode_daz(
1004 ; GFX90A-SAME: ptr addrspace(1) [[PTR:%.*]], <2 x half> [[VALUE:%.*]]) #[[ATTR1]] {
1005 ; GFX90A-NEXT: [[RES:%.*]] = atomicrmw fadd ptr addrspace(1) [[PTR]], <2 x half> [[VALUE]] seq_cst, align 4, !amdgpu.no.fine.grained.memory [[META0]], !amdgpu.no.remote.memory [[META0]], !amdgpu.ignore.denormal.mode [[META0]]
1006 ; GFX90A-NEXT: ret <2 x half> [[RES]]
1007 ;
1008 ; GFX940-LABEL: define <2 x half> @test_atomicrmw_fadd_v2f16_global_system__amdgpu_ignore_denormal_mode__amdgpu_no_fine_grained_memory__amdgpu_no_remote_memory__denormal_mode_daz(
1009 ; GFX940-SAME: ptr addrspace(1) [[PTR:%.*]], <2 x half> [[VALUE:%.*]]) #[[ATTR1]] {
1010 ; GFX940-NEXT: [[RES:%.*]] = atomicrmw fadd ptr addrspace(1) [[PTR]], <2 x half> [[VALUE]] seq_cst, align 4, !amdgpu.no.fine.grained.memory [[META0]], !amdgpu.no.remote.memory [[META0]], !amdgpu.ignore.denormal.mode [[META0]]
1011 ; GFX940-NEXT: ret <2 x half> [[RES]]
1012 ;
1013 ; GFX10-LABEL: define <2 x half> @test_atomicrmw_fadd_v2f16_global_system__amdgpu_ignore_denormal_mode__amdgpu_no_fine_grained_memory__amdgpu_no_remote_memory__denormal_mode_daz(
1014 ; GFX10-SAME: ptr addrspace(1) [[PTR:%.*]], <2 x half> [[VALUE:%.*]]) #[[ATTR1]] {
1015 ; GFX10-NEXT: [[TMP1:%.*]] = load <2 x half>, ptr addrspace(1) [[PTR]], align 4
1016 ; GFX10-NEXT: br label [[ATOMICRMW_START:%.*]]
1017 ; GFX10: atomicrmw.start:
1018 ; GFX10-NEXT: [[LOADED:%.*]] = phi <2 x half> [ [[TMP1]], [[TMP0:%.*]] ], [ [[TMP5:%.*]], [[ATOMICRMW_START]] ]
1019 ; GFX10-NEXT: [[NEW:%.*]] = fadd <2 x half> [[LOADED]], [[VALUE]]
1020 ; GFX10-NEXT: [[TMP2:%.*]] = bitcast <2 x half> [[NEW]] to i32
1021 ; GFX10-NEXT: [[TMP3:%.*]] = bitcast <2 x half> [[LOADED]] to i32
1022 ; GFX10-NEXT: [[TMP4:%.*]] = cmpxchg ptr addrspace(1) [[PTR]], i32 [[TMP3]], i32 [[TMP2]] seq_cst seq_cst, align 4, !amdgpu.no.fine.grained.memory [[META0]], !amdgpu.no.remote.memory [[META0]]
1023 ; GFX10-NEXT: [[SUCCESS:%.*]] = extractvalue { i32, i1 } [[TMP4]], 1
1024 ; GFX10-NEXT: [[NEWLOADED:%.*]] = extractvalue { i32, i1 } [[TMP4]], 0
1025 ; GFX10-NEXT: [[TMP5]] = bitcast i32 [[NEWLOADED]] to <2 x half>
1026 ; GFX10-NEXT: br i1 [[SUCCESS]], label [[ATOMICRMW_END:%.*]], label [[ATOMICRMW_START]]
1027 ; GFX10: atomicrmw.end:
1028 ; GFX10-NEXT: ret <2 x half> [[TMP5]]
1029 ;
1030 ; GFX11-LABEL: define <2 x half> @test_atomicrmw_fadd_v2f16_global_system__amdgpu_ignore_denormal_mode__amdgpu_no_fine_grained_memory__amdgpu_no_remote_memory__denormal_mode_daz(
1031 ; GFX11-SAME: ptr addrspace(1) [[PTR:%.*]], <2 x half> [[VALUE:%.*]]) #[[ATTR1]] {
1032 ; GFX11-NEXT: [[TMP1:%.*]] = load <2 x half>, ptr addrspace(1) [[PTR]], align 4
1033 ; GFX11-NEXT: br label [[ATOMICRMW_START:%.*]]
1034 ; GFX11: atomicrmw.start:
1035 ; GFX11-NEXT: [[LOADED:%.*]] = phi <2 x half> [ [[TMP1]], [[TMP0:%.*]] ], [ [[TMP5:%.*]], [[ATOMICRMW_START]] ]
1036 ; GFX11-NEXT: [[NEW:%.*]] = fadd <2 x half> [[LOADED]], [[VALUE]]
1037 ; GFX11-NEXT: [[TMP2:%.*]] = bitcast <2 x half> [[NEW]] to i32
1038 ; GFX11-NEXT: [[TMP3:%.*]] = bitcast <2 x half> [[LOADED]] to i32
1039 ; GFX11-NEXT: [[TMP4:%.*]] = cmpxchg ptr addrspace(1) [[PTR]], i32 [[TMP3]], i32 [[TMP2]] seq_cst seq_cst, align 4, !amdgpu.no.fine.grained.memory [[META0]], !amdgpu.no.remote.memory [[META0]]
1040 ; GFX11-NEXT: [[SUCCESS:%.*]] = extractvalue { i32, i1 } [[TMP4]], 1
1041 ; GFX11-NEXT: [[NEWLOADED:%.*]] = extractvalue { i32, i1 } [[TMP4]], 0
1042 ; GFX11-NEXT: [[TMP5]] = bitcast i32 [[NEWLOADED]] to <2 x half>
1043 ; GFX11-NEXT: br i1 [[SUCCESS]], label [[ATOMICRMW_END:%.*]], label [[ATOMICRMW_START]]
1044 ; GFX11: atomicrmw.end:
1045 ; GFX11-NEXT: ret <2 x half> [[TMP5]]
1046 ;
1047 ; GFX12-LABEL: define <2 x half> @test_atomicrmw_fadd_v2f16_global_system__amdgpu_ignore_denormal_mode__amdgpu_no_fine_grained_memory__amdgpu_no_remote_memory__denormal_mode_daz(
1048 ; GFX12-SAME: ptr addrspace(1) [[PTR:%.*]], <2 x half> [[VALUE:%.*]]) #[[ATTR1]] {
1049 ; GFX12-NEXT: [[RES:%.*]] = atomicrmw fadd ptr addrspace(1) [[PTR]], <2 x half> [[VALUE]] seq_cst, align 4, !amdgpu.no.fine.grained.memory [[META0]], !amdgpu.no.remote.memory [[META0]], !amdgpu.ignore.denormal.mode [[META0]]
1050 ; GFX12-NEXT: ret <2 x half> [[RES]]
1051 ;
1052 %res = atomicrmw fadd ptr addrspace(1) %ptr, <2 x half> %value seq_cst, align 4, !amdgpu.no.fine.grained.memory !0, !amdgpu.no.remote.memory !0, !amdgpu.ignore.denormal.mode !0
1053 ret <2 x half> %res
1054 }
1056 define <2 x half> @test_atomicrmw_fadd_v2f16_global_system__amdgpu_ignore_denormal_mode__amdgpu_no_fine_grained_memory__amdgpu_no_remote_memory__denormal_mode_dynamic(ptr addrspace(1) %ptr, <2 x half> %value) #1 {
1057 ; GFX803-LABEL: define <2 x half> @test_atomicrmw_fadd_v2f16_global_system__amdgpu_ignore_denormal_mode__amdgpu_no_fine_grained_memory__amdgpu_no_remote_memory__denormal_mode_dynamic(
1058 ; GFX803-SAME: ptr addrspace(1) [[PTR:%.*]], <2 x half> [[VALUE:%.*]]) #[[ATTR2]] {
1059 ; GFX803-NEXT: [[TMP1:%.*]] = load <2 x half>, ptr addrspace(1) [[PTR]], align 4
1060 ; GFX803-NEXT: br label [[ATOMICRMW_START:%.*]]
1061 ; GFX803: atomicrmw.start:
1062 ; GFX803-NEXT: [[LOADED:%.*]] = phi <2 x half> [ [[TMP1]], [[TMP0:%.*]] ], [ [[TMP5:%.*]], [[ATOMICRMW_START]] ]
1063 ; GFX803-NEXT: [[NEW:%.*]] = fadd <2 x half> [[LOADED]], [[VALUE]]
1064 ; GFX803-NEXT: [[TMP2:%.*]] = bitcast <2 x half> [[NEW]] to i32
1065 ; GFX803-NEXT: [[TMP3:%.*]] = bitcast <2 x half> [[LOADED]] to i32
1066 ; GFX803-NEXT: [[TMP4:%.*]] = cmpxchg ptr addrspace(1) [[PTR]], i32 [[TMP3]], i32 [[TMP2]] seq_cst seq_cst, align 4, !amdgpu.no.fine.grained.memory [[META0]], !amdgpu.no.remote.memory [[META0]]
1067 ; GFX803-NEXT: [[SUCCESS:%.*]] = extractvalue { i32, i1 } [[TMP4]], 1
1068 ; GFX803-NEXT: [[NEWLOADED:%.*]] = extractvalue { i32, i1 } [[TMP4]], 0
1069 ; GFX803-NEXT: [[TMP5]] = bitcast i32 [[NEWLOADED]] to <2 x half>
1070 ; GFX803-NEXT: br i1 [[SUCCESS]], label [[ATOMICRMW_END:%.*]], label [[ATOMICRMW_START]]
1071 ; GFX803: atomicrmw.end:
1072 ; GFX803-NEXT: ret <2 x half> [[TMP5]]
1073 ;
1074 ; GFX906-LABEL: define <2 x half> @test_atomicrmw_fadd_v2f16_global_system__amdgpu_ignore_denormal_mode__amdgpu_no_fine_grained_memory__amdgpu_no_remote_memory__denormal_mode_dynamic(
1075 ; GFX906-SAME: ptr addrspace(1) [[PTR:%.*]], <2 x half> [[VALUE:%.*]]) #[[ATTR2]] {
1076 ; GFX906-NEXT: [[TMP1:%.*]] = load <2 x half>, ptr addrspace(1) [[PTR]], align 4
1077 ; GFX906-NEXT: br label [[ATOMICRMW_START:%.*]]
1078 ; GFX906: atomicrmw.start:
1079 ; GFX906-NEXT: [[LOADED:%.*]] = phi <2 x half> [ [[TMP1]], [[TMP0:%.*]] ], [ [[TMP5:%.*]], [[ATOMICRMW_START]] ]
1080 ; GFX906-NEXT: [[NEW:%.*]] = fadd <2 x half> [[LOADED]], [[VALUE]]
1081 ; GFX906-NEXT: [[TMP2:%.*]] = bitcast <2 x half> [[NEW]] to i32
1082 ; GFX906-NEXT: [[TMP3:%.*]] = bitcast <2 x half> [[LOADED]] to i32
1083 ; GFX906-NEXT: [[TMP4:%.*]] = cmpxchg ptr addrspace(1) [[PTR]], i32 [[TMP3]], i32 [[TMP2]] seq_cst seq_cst, align 4, !amdgpu.no.fine.grained.memory [[META0]], !amdgpu.no.remote.memory [[META0]]
1084 ; GFX906-NEXT: [[SUCCESS:%.*]] = extractvalue { i32, i1 } [[TMP4]], 1
1085 ; GFX906-NEXT: [[NEWLOADED:%.*]] = extractvalue { i32, i1 } [[TMP4]], 0
1086 ; GFX906-NEXT: [[TMP5]] = bitcast i32 [[NEWLOADED]] to <2 x half>
1087 ; GFX906-NEXT: br i1 [[SUCCESS]], label [[ATOMICRMW_END:%.*]], label [[ATOMICRMW_START]]
1088 ; GFX906: atomicrmw.end:
1089 ; GFX906-NEXT: ret <2 x half> [[TMP5]]
1090 ;
1091 ; GFX908-LABEL: define <2 x half> @test_atomicrmw_fadd_v2f16_global_system__amdgpu_ignore_denormal_mode__amdgpu_no_fine_grained_memory__amdgpu_no_remote_memory__denormal_mode_dynamic(
1092 ; GFX908-SAME: ptr addrspace(1) [[PTR:%.*]], <2 x half> [[VALUE:%.*]]) #[[ATTR2]] {
1093 ; GFX908-NEXT: [[TMP1:%.*]] = load <2 x half>, ptr addrspace(1) [[PTR]], align 4
1094 ; GFX908-NEXT: br label [[ATOMICRMW_START:%.*]]
1095 ; GFX908: atomicrmw.start:
1096 ; GFX908-NEXT: [[LOADED:%.*]] = phi <2 x half> [ [[TMP1]], [[TMP0:%.*]] ], [ [[TMP5:%.*]], [[ATOMICRMW_START]] ]
1097 ; GFX908-NEXT: [[NEW:%.*]] = fadd <2 x half> [[LOADED]], [[VALUE]]
1098 ; GFX908-NEXT: [[TMP2:%.*]] = bitcast <2 x half> [[NEW]] to i32
1099 ; GFX908-NEXT: [[TMP3:%.*]] = bitcast <2 x half> [[LOADED]] to i32
1100 ; GFX908-NEXT: [[TMP4:%.*]] = cmpxchg ptr addrspace(1) [[PTR]], i32 [[TMP3]], i32 [[TMP2]] seq_cst seq_cst, align 4, !amdgpu.no.fine.grained.memory [[META0]], !amdgpu.no.remote.memory [[META0]]
1101 ; GFX908-NEXT: [[SUCCESS:%.*]] = extractvalue { i32, i1 } [[TMP4]], 1
1102 ; GFX908-NEXT: [[NEWLOADED:%.*]] = extractvalue { i32, i1 } [[TMP4]], 0
1103 ; GFX908-NEXT: [[TMP5]] = bitcast i32 [[NEWLOADED]] to <2 x half>
1104 ; GFX908-NEXT: br i1 [[SUCCESS]], label [[ATOMICRMW_END:%.*]], label [[ATOMICRMW_START]]
1105 ; GFX908: atomicrmw.end:
1106 ; GFX908-NEXT: ret <2 x half> [[TMP5]]
1107 ;
1108 ; GFX90A-LABEL: define <2 x half> @test_atomicrmw_fadd_v2f16_global_system__amdgpu_ignore_denormal_mode__amdgpu_no_fine_grained_memory__amdgpu_no_remote_memory__denormal_mode_dynamic(
1109 ; GFX90A-SAME: ptr addrspace(1) [[PTR:%.*]], <2 x half> [[VALUE:%.*]]) #[[ATTR2]] {
1110 ; GFX90A-NEXT: [[RES:%.*]] = atomicrmw fadd ptr addrspace(1) [[PTR]], <2 x half> [[VALUE]] seq_cst, align 4, !amdgpu.no.fine.grained.memory [[META0]], !amdgpu.no.remote.memory [[META0]], !amdgpu.ignore.denormal.mode [[META0]]
1111 ; GFX90A-NEXT: ret <2 x half> [[RES]]
1112 ;
1113 ; GFX940-LABEL: define <2 x half> @test_atomicrmw_fadd_v2f16_global_system__amdgpu_ignore_denormal_mode__amdgpu_no_fine_grained_memory__amdgpu_no_remote_memory__denormal_mode_dynamic(
1114 ; GFX940-SAME: ptr addrspace(1) [[PTR:%.*]], <2 x half> [[VALUE:%.*]]) #[[ATTR2]] {
1115 ; GFX940-NEXT: [[RES:%.*]] = atomicrmw fadd ptr addrspace(1) [[PTR]], <2 x half> [[VALUE]] seq_cst, align 4, !amdgpu.no.fine.grained.memory [[META0]], !amdgpu.no.remote.memory [[META0]], !amdgpu.ignore.denormal.mode [[META0]]
1116 ; GFX940-NEXT: ret <2 x half> [[RES]]
1117 ;
1118 ; GFX10-LABEL: define <2 x half> @test_atomicrmw_fadd_v2f16_global_system__amdgpu_ignore_denormal_mode__amdgpu_no_fine_grained_memory__amdgpu_no_remote_memory__denormal_mode_dynamic(
1119 ; GFX10-SAME: ptr addrspace(1) [[PTR:%.*]], <2 x half> [[VALUE:%.*]]) #[[ATTR2]] {
1120 ; GFX10-NEXT: [[TMP1:%.*]] = load <2 x half>, ptr addrspace(1) [[PTR]], align 4
1121 ; GFX10-NEXT: br label [[ATOMICRMW_START:%.*]]
1122 ; GFX10: atomicrmw.start:
1123 ; GFX10-NEXT: [[LOADED:%.*]] = phi <2 x half> [ [[TMP1]], [[TMP0:%.*]] ], [ [[TMP5:%.*]], [[ATOMICRMW_START]] ]
1124 ; GFX10-NEXT: [[NEW:%.*]] = fadd <2 x half> [[LOADED]], [[VALUE]]
1125 ; GFX10-NEXT: [[TMP2:%.*]] = bitcast <2 x half> [[NEW]] to i32
1126 ; GFX10-NEXT: [[TMP3:%.*]] = bitcast <2 x half> [[LOADED]] to i32
1127 ; GFX10-NEXT: [[TMP4:%.*]] = cmpxchg ptr addrspace(1) [[PTR]], i32 [[TMP3]], i32 [[TMP2]] seq_cst seq_cst, align 4, !amdgpu.no.fine.grained.memory [[META0]], !amdgpu.no.remote.memory [[META0]]
1128 ; GFX10-NEXT: [[SUCCESS:%.*]] = extractvalue { i32, i1 } [[TMP4]], 1
1129 ; GFX10-NEXT: [[NEWLOADED:%.*]] = extractvalue { i32, i1 } [[TMP4]], 0
1130 ; GFX10-NEXT: [[TMP5]] = bitcast i32 [[NEWLOADED]] to <2 x half>
1131 ; GFX10-NEXT: br i1 [[SUCCESS]], label [[ATOMICRMW_END:%.*]], label [[ATOMICRMW_START]]
1132 ; GFX10: atomicrmw.end:
1133 ; GFX10-NEXT: ret <2 x half> [[TMP5]]
1134 ;
1135 ; GFX11-LABEL: define <2 x half> @test_atomicrmw_fadd_v2f16_global_system__amdgpu_ignore_denormal_mode__amdgpu_no_fine_grained_memory__amdgpu_no_remote_memory__denormal_mode_dynamic(
1136 ; GFX11-SAME: ptr addrspace(1) [[PTR:%.*]], <2 x half> [[VALUE:%.*]]) #[[ATTR2]] {
1137 ; GFX11-NEXT: [[TMP1:%.*]] = load <2 x half>, ptr addrspace(1) [[PTR]], align 4
1138 ; GFX11-NEXT: br label [[ATOMICRMW_START:%.*]]
1139 ; GFX11: atomicrmw.start:
1140 ; GFX11-NEXT: [[LOADED:%.*]] = phi <2 x half> [ [[TMP1]], [[TMP0:%.*]] ], [ [[TMP5:%.*]], [[ATOMICRMW_START]] ]
1141 ; GFX11-NEXT: [[NEW:%.*]] = fadd <2 x half> [[LOADED]], [[VALUE]]
1142 ; GFX11-NEXT: [[TMP2:%.*]] = bitcast <2 x half> [[NEW]] to i32
1143 ; GFX11-NEXT: [[TMP3:%.*]] = bitcast <2 x half> [[LOADED]] to i32
1144 ; GFX11-NEXT: [[TMP4:%.*]] = cmpxchg ptr addrspace(1) [[PTR]], i32 [[TMP3]], i32 [[TMP2]] seq_cst seq_cst, align 4, !amdgpu.no.fine.grained.memory [[META0]], !amdgpu.no.remote.memory [[META0]]
1145 ; GFX11-NEXT: [[SUCCESS:%.*]] = extractvalue { i32, i1 } [[TMP4]], 1
1146 ; GFX11-NEXT: [[NEWLOADED:%.*]] = extractvalue { i32, i1 } [[TMP4]], 0
1147 ; GFX11-NEXT: [[TMP5]] = bitcast i32 [[NEWLOADED]] to <2 x half>
1148 ; GFX11-NEXT: br i1 [[SUCCESS]], label [[ATOMICRMW_END:%.*]], label [[ATOMICRMW_START]]
1149 ; GFX11: atomicrmw.end:
1150 ; GFX11-NEXT: ret <2 x half> [[TMP5]]
1151 ;
1152 ; GFX12-LABEL: define <2 x half> @test_atomicrmw_fadd_v2f16_global_system__amdgpu_ignore_denormal_mode__amdgpu_no_fine_grained_memory__amdgpu_no_remote_memory__denormal_mode_dynamic(
1153 ; GFX12-SAME: ptr addrspace(1) [[PTR:%.*]], <2 x half> [[VALUE:%.*]]) #[[ATTR2]] {
1154 ; GFX12-NEXT: [[RES:%.*]] = atomicrmw fadd ptr addrspace(1) [[PTR]], <2 x half> [[VALUE]] seq_cst, align 4, !amdgpu.no.fine.grained.memory [[META0]], !amdgpu.no.remote.memory [[META0]], !amdgpu.ignore.denormal.mode [[META0]]
1155 ; GFX12-NEXT: ret <2 x half> [[RES]]
1156 ;
1157 %res = atomicrmw fadd ptr addrspace(1) %ptr, <2 x half> %value seq_cst, align 4, !amdgpu.no.fine.grained.memory !0, !amdgpu.no.remote.memory !0, !amdgpu.ignore.denormal.mode !0
1158 ret <2 x half> %res
1159 }
1161 ;---------------------------------------------------------------------
1162 ; atomicrmw fsub
1163 ;---------------------------------------------------------------------
1165 define <2 x half> @test_atomicrmw_fsub_v2f16_global_system(ptr addrspace(1) %ptr, <2 x half> %value) {
1166 ; COMMON-LABEL: define <2 x half> @test_atomicrmw_fsub_v2f16_global_system(
1167 ; COMMON-SAME: ptr addrspace(1) [[PTR:%.*]], <2 x half> [[VALUE:%.*]]) #[[ATTR0]] {
1168 ; COMMON-NEXT: [[TMP1:%.*]] = load <2 x half>, ptr addrspace(1) [[PTR]], align 4
1169 ; COMMON-NEXT: br label [[ATOMICRMW_START:%.*]]
1170 ; COMMON: atomicrmw.start:
1171 ; COMMON-NEXT: [[LOADED:%.*]] = phi <2 x half> [ [[TMP1]], [[TMP0:%.*]] ], [ [[RES:%.*]], [[ATOMICRMW_START]] ]
1172 ; COMMON-NEXT: [[NEW:%.*]] = fsub <2 x half> [[LOADED]], [[VALUE]]
1173 ; COMMON-NEXT: [[TMP2:%.*]] = bitcast <2 x half> [[NEW]] to i32
1174 ; COMMON-NEXT: [[TMP3:%.*]] = bitcast <2 x half> [[LOADED]] to i32
1175 ; COMMON-NEXT: [[TMP4:%.*]] = cmpxchg ptr addrspace(1) [[PTR]], i32 [[TMP3]], i32 [[TMP2]] seq_cst seq_cst, align 4
1176 ; COMMON-NEXT: [[SUCCESS:%.*]] = extractvalue { i32, i1 } [[TMP4]], 1
1177 ; COMMON-NEXT: [[NEWLOADED:%.*]] = extractvalue { i32, i1 } [[TMP4]], 0
1178 ; COMMON-NEXT: [[RES]] = bitcast i32 [[NEWLOADED]] to <2 x half>
1179 ; COMMON-NEXT: br i1 [[SUCCESS]], label [[ATOMICRMW_END:%.*]], label [[ATOMICRMW_START]]
1180 ; COMMON: atomicrmw.end:
1181 ; COMMON-NEXT: ret <2 x half> [[RES]]
1182 ;
1183 %res = atomicrmw fsub ptr addrspace(1) %ptr, <2 x half> %value seq_cst
1184 ret <2 x half> %res
1185 }
1187 define <2 x half> @test_atomicrmw_fsub_v2f16_global_system__amdgpu_no_fine_grained_memory(ptr addrspace(1) %ptr, <2 x half> %value) {
1188 ; COMMON-LABEL: define <2 x half> @test_atomicrmw_fsub_v2f16_global_system__amdgpu_no_fine_grained_memory(
1189 ; COMMON-SAME: ptr addrspace(1) [[PTR:%.*]], <2 x half> [[VALUE:%.*]]) #[[ATTR0]] {
1190 ; COMMON-NEXT: [[TMP1:%.*]] = load <2 x half>, ptr addrspace(1) [[PTR]], align 4
1191 ; COMMON-NEXT: br label [[ATOMICRMW_START:%.*]]
1192 ; COMMON: atomicrmw.start:
1193 ; COMMON-NEXT: [[LOADED:%.*]] = phi <2 x half> [ [[TMP1]], [[TMP0:%.*]] ], [ [[RES:%.*]], [[ATOMICRMW_START]] ]
1194 ; COMMON-NEXT: [[NEW:%.*]] = fsub <2 x half> [[LOADED]], [[VALUE]]
1195 ; COMMON-NEXT: [[TMP2:%.*]] = bitcast <2 x half> [[NEW]] to i32
1196 ; COMMON-NEXT: [[TMP3:%.*]] = bitcast <2 x half> [[LOADED]] to i32
1197 ; COMMON-NEXT: [[TMP4:%.*]] = cmpxchg ptr addrspace(1) [[PTR]], i32 [[TMP3]], i32 [[TMP2]] seq_cst seq_cst, align 4, !amdgpu.no.fine.grained.memory [[META0:![0-9]+]]
1198 ; COMMON-NEXT: [[SUCCESS:%.*]] = extractvalue { i32, i1 } [[TMP4]], 1
1199 ; COMMON-NEXT: [[NEWLOADED:%.*]] = extractvalue { i32, i1 } [[TMP4]], 0
1200 ; COMMON-NEXT: [[RES]] = bitcast i32 [[NEWLOADED]] to <2 x half>
1201 ; COMMON-NEXT: br i1 [[SUCCESS]], label [[ATOMICRMW_END:%.*]], label [[ATOMICRMW_START]]
1202 ; COMMON: atomicrmw.end:
1203 ; COMMON-NEXT: ret <2 x half> [[RES]]
1204 ;
1205 %res = atomicrmw fsub ptr addrspace(1) %ptr, <2 x half> %value seq_cst, !amdgpu.no.fine.grained.memory !0
1206 ret <2 x half> %res
1207 }
1209 define <2 x half> @test_atomicrmw_fsub_v2f16_global_system__amdgpu_no_remote_memory(ptr addrspace(1) %ptr, <2 x half> %value) {
1210 ; COMMON-LABEL: define <2 x half> @test_atomicrmw_fsub_v2f16_global_system__amdgpu_no_remote_memory(
1211 ; COMMON-SAME: ptr addrspace(1) [[PTR:%.*]], <2 x half> [[VALUE:%.*]]) #[[ATTR0]] {
1212 ; COMMON-NEXT: [[TMP1:%.*]] = load <2 x half>, ptr addrspace(1) [[PTR]], align 4
1213 ; COMMON-NEXT: br label [[ATOMICRMW_START:%.*]]
1214 ; COMMON: atomicrmw.start:
1215 ; COMMON-NEXT: [[LOADED:%.*]] = phi <2 x half> [ [[TMP1]], [[TMP0:%.*]] ], [ [[RES:%.*]], [[ATOMICRMW_START]] ]
1216 ; COMMON-NEXT: [[NEW:%.*]] = fsub <2 x half> [[LOADED]], [[VALUE]]
1217 ; COMMON-NEXT: [[TMP2:%.*]] = bitcast <2 x half> [[NEW]] to i32
1218 ; COMMON-NEXT: [[TMP3:%.*]] = bitcast <2 x half> [[LOADED]] to i32
1219 ; COMMON-NEXT: [[TMP4:%.*]] = cmpxchg ptr addrspace(1) [[PTR]], i32 [[TMP3]], i32 [[TMP2]] seq_cst seq_cst, align 4, !amdgpu.no.remote.memory [[META0]]
1220 ; COMMON-NEXT: [[SUCCESS:%.*]] = extractvalue { i32, i1 } [[TMP4]], 1
1221 ; COMMON-NEXT: [[NEWLOADED:%.*]] = extractvalue { i32, i1 } [[TMP4]], 0
1222 ; COMMON-NEXT: [[RES]] = bitcast i32 [[NEWLOADED]] to <2 x half>
1223 ; COMMON-NEXT: br i1 [[SUCCESS]], label [[ATOMICRMW_END:%.*]], label [[ATOMICRMW_START]]
1224 ; COMMON: atomicrmw.end:
1225 ; COMMON-NEXT: ret <2 x half> [[RES]]
1226 ;
1227 %res = atomicrmw fsub ptr addrspace(1) %ptr, <2 x half> %value seq_cst, !amdgpu.no.remote.memory !0
1228 ret <2 x half> %res
1229 }
1231 define <2 x half> @test_atomicrmw_fsub_v2f16_global_system__amdgpu_no_fine_grained_memory__amdgpu_no_remote_memory(ptr addrspace(1) %ptr, <2 x half> %value) {
1232 ; COMMON-LABEL: define <2 x half> @test_atomicrmw_fsub_v2f16_global_system__amdgpu_no_fine_grained_memory__amdgpu_no_remote_memory(
1233 ; COMMON-SAME: ptr addrspace(1) [[PTR:%.*]], <2 x half> [[VALUE:%.*]]) #[[ATTR0]] {
1234 ; COMMON-NEXT: [[TMP1:%.*]] = load <2 x half>, ptr addrspace(1) [[PTR]], align 4
1235 ; COMMON-NEXT: br label [[ATOMICRMW_START:%.*]]
1236 ; COMMON: atomicrmw.start:
1237 ; COMMON-NEXT: [[LOADED:%.*]] = phi <2 x half> [ [[TMP1]], [[TMP0:%.*]] ], [ [[RES:%.*]], [[ATOMICRMW_START]] ]
1238 ; COMMON-NEXT: [[NEW:%.*]] = fsub <2 x half> [[LOADED]], [[VALUE]]
1239 ; COMMON-NEXT: [[TMP2:%.*]] = bitcast <2 x half> [[NEW]] to i32
1240 ; COMMON-NEXT: [[TMP3:%.*]] = bitcast <2 x half> [[LOADED]] to i32
1241 ; COMMON-NEXT: [[TMP4:%.*]] = cmpxchg ptr addrspace(1) [[PTR]], i32 [[TMP3]], i32 [[TMP2]] seq_cst seq_cst, align 4, !amdgpu.no.fine.grained.memory [[META0]], !amdgpu.no.remote.memory [[META0]]
1242 ; COMMON-NEXT: [[SUCCESS:%.*]] = extractvalue { i32, i1 } [[TMP4]], 1
1243 ; COMMON-NEXT: [[NEWLOADED:%.*]] = extractvalue { i32, i1 } [[TMP4]], 0
1244 ; COMMON-NEXT: [[RES]] = bitcast i32 [[NEWLOADED]] to <2 x half>
1245 ; COMMON-NEXT: br i1 [[SUCCESS]], label [[ATOMICRMW_END:%.*]], label [[ATOMICRMW_START]]
1246 ; COMMON: atomicrmw.end:
1247 ; COMMON-NEXT: ret <2 x half> [[RES]]
1248 ;
1249 %res = atomicrmw fsub ptr addrspace(1) %ptr, <2 x half> %value seq_cst, !amdgpu.no.fine.grained.memory !0, !amdgpu.no.remote.memory !0
1250 ret <2 x half> %res
1251 }
1253 define <2 x half> @test_atomicrmw_fsub_v2f16_global_system__amdgpu_ignore_denormal_mode(ptr addrspace(1) %ptr, <2 x half> %value) {
1254 ; COMMON-LABEL: define <2 x half> @test_atomicrmw_fsub_v2f16_global_system__amdgpu_ignore_denormal_mode(
1255 ; COMMON-SAME: ptr addrspace(1) [[PTR:%.*]], <2 x half> [[VALUE:%.*]]) #[[ATTR0]] {
1256 ; COMMON-NEXT: [[TMP1:%.*]] = load <2 x half>, ptr addrspace(1) [[PTR]], align 4
1257 ; COMMON-NEXT: br label [[ATOMICRMW_START:%.*]]
1258 ; COMMON: atomicrmw.start:
1259 ; COMMON-NEXT: [[LOADED:%.*]] = phi <2 x half> [ [[TMP1]], [[TMP0:%.*]] ], [ [[TMP5:%.*]], [[ATOMICRMW_START]] ]
1260 ; COMMON-NEXT: [[NEW:%.*]] = fsub <2 x half> [[LOADED]], [[VALUE]]
1261 ; COMMON-NEXT: [[TMP2:%.*]] = bitcast <2 x half> [[NEW]] to i32
1262 ; COMMON-NEXT: [[TMP3:%.*]] = bitcast <2 x half> [[LOADED]] to i32
1263 ; COMMON-NEXT: [[TMP4:%.*]] = cmpxchg ptr addrspace(1) [[PTR]], i32 [[TMP3]], i32 [[TMP2]] seq_cst seq_cst, align 4
1264 ; COMMON-NEXT: [[SUCCESS:%.*]] = extractvalue { i32, i1 } [[TMP4]], 1
1265 ; COMMON-NEXT: [[NEWLOADED:%.*]] = extractvalue { i32, i1 } [[TMP4]], 0
1266 ; COMMON-NEXT: [[TMP5]] = bitcast i32 [[NEWLOADED]] to <2 x half>
1267 ; COMMON-NEXT: br i1 [[SUCCESS]], label [[ATOMICRMW_END:%.*]], label [[ATOMICRMW_START]]
1268 ; COMMON: atomicrmw.end:
1269 ; COMMON-NEXT: ret <2 x half> [[TMP5]]
1270 ;
1271 %res = atomicrmw fsub ptr addrspace(1) %ptr, <2 x half> %value seq_cst, align 4, !amdgpu.ignore.denormal.mode !0
1272 ret <2 x half> %res
1273 }
1275 define <2 x half> @test_atomicrmw_fsub_v2f16_global_system__amdgpu_ignore_denormal_mode__amdgpu_no_fine_grained_memory(ptr addrspace(1) %ptr, <2 x half> %value) {
1276 ; COMMON-LABEL: define <2 x half> @test_atomicrmw_fsub_v2f16_global_system__amdgpu_ignore_denormal_mode__amdgpu_no_fine_grained_memory(
1277 ; COMMON-SAME: ptr addrspace(1) [[PTR:%.*]], <2 x half> [[VALUE:%.*]]) #[[ATTR0]] {
1278 ; COMMON-NEXT: [[TMP1:%.*]] = load <2 x half>, ptr addrspace(1) [[PTR]], align 4
1279 ; COMMON-NEXT: br label [[ATOMICRMW_START:%.*]]
1280 ; COMMON: atomicrmw.start:
1281 ; COMMON-NEXT: [[LOADED:%.*]] = phi <2 x half> [ [[TMP1]], [[TMP0:%.*]] ], [ [[TMP5:%.*]], [[ATOMICRMW_START]] ]
1282 ; COMMON-NEXT: [[NEW:%.*]] = fsub <2 x half> [[LOADED]], [[VALUE]]
1283 ; COMMON-NEXT: [[TMP2:%.*]] = bitcast <2 x half> [[NEW]] to i32
1284 ; COMMON-NEXT: [[TMP3:%.*]] = bitcast <2 x half> [[LOADED]] to i32
1285 ; COMMON-NEXT: [[TMP4:%.*]] = cmpxchg ptr addrspace(1) [[PTR]], i32 [[TMP3]], i32 [[TMP2]] seq_cst seq_cst, align 4, !amdgpu.no.fine.grained.memory [[META0]]
1286 ; COMMON-NEXT: [[SUCCESS:%.*]] = extractvalue { i32, i1 } [[TMP4]], 1
1287 ; COMMON-NEXT: [[NEWLOADED:%.*]] = extractvalue { i32, i1 } [[TMP4]], 0
1288 ; COMMON-NEXT: [[TMP5]] = bitcast i32 [[NEWLOADED]] to <2 x half>
1289 ; COMMON-NEXT: br i1 [[SUCCESS]], label [[ATOMICRMW_END:%.*]], label [[ATOMICRMW_START]]
1290 ; COMMON: atomicrmw.end:
1291 ; COMMON-NEXT: ret <2 x half> [[TMP5]]
1292 ;
1293 %res = atomicrmw fsub ptr addrspace(1) %ptr, <2 x half> %value seq_cst, align 4, !amdgpu.no.fine.grained.memory !0, !amdgpu.ignore.denormal.mode !0
1294 ret <2 x half> %res
1295 }
1297 define <2 x half> @test_atomicrmw_fsub_v2f16_global_system__amdgpu_ignore_denormal_mode__amdgpu_no_remote_memory(ptr addrspace(1) %ptr, <2 x half> %value) {
1298 ; COMMON-LABEL: define <2 x half> @test_atomicrmw_fsub_v2f16_global_system__amdgpu_ignore_denormal_mode__amdgpu_no_remote_memory(
1299 ; COMMON-SAME: ptr addrspace(1) [[PTR:%.*]], <2 x half> [[VALUE:%.*]]) #[[ATTR0]] {
1300 ; COMMON-NEXT: [[TMP1:%.*]] = load <2 x half>, ptr addrspace(1) [[PTR]], align 4
1301 ; COMMON-NEXT: br label [[ATOMICRMW_START:%.*]]
1302 ; COMMON: atomicrmw.start:
1303 ; COMMON-NEXT: [[LOADED:%.*]] = phi <2 x half> [ [[TMP1]], [[TMP0:%.*]] ], [ [[TMP5:%.*]], [[ATOMICRMW_START]] ]
1304 ; COMMON-NEXT: [[NEW:%.*]] = fsub <2 x half> [[LOADED]], [[VALUE]]
1305 ; COMMON-NEXT: [[TMP2:%.*]] = bitcast <2 x half> [[NEW]] to i32
1306 ; COMMON-NEXT: [[TMP3:%.*]] = bitcast <2 x half> [[LOADED]] to i32
1307 ; COMMON-NEXT: [[TMP4:%.*]] = cmpxchg ptr addrspace(1) [[PTR]], i32 [[TMP3]], i32 [[TMP2]] seq_cst seq_cst, align 4, !amdgpu.no.remote.memory [[META0]]
1308 ; COMMON-NEXT: [[SUCCESS:%.*]] = extractvalue { i32, i1 } [[TMP4]], 1
1309 ; COMMON-NEXT: [[NEWLOADED:%.*]] = extractvalue { i32, i1 } [[TMP4]], 0
1310 ; COMMON-NEXT: [[TMP5]] = bitcast i32 [[NEWLOADED]] to <2 x half>
1311 ; COMMON-NEXT: br i1 [[SUCCESS]], label [[ATOMICRMW_END:%.*]], label [[ATOMICRMW_START]]
1312 ; COMMON: atomicrmw.end:
1313 ; COMMON-NEXT: ret <2 x half> [[TMP5]]
1314 ;
1315 %res = atomicrmw fsub ptr addrspace(1) %ptr, <2 x half> %value seq_cst, align 4, !amdgpu.no.remote.memory !0, !amdgpu.ignore.denormal.mode !0
1316 ret <2 x half> %res
1317 }
1319 define <2 x half> @test_atomicrmw_fsub_v2f16_global_system__amdgpu_ignore_denormal_mode__amdgpu_no_fine_grained_memory__amdgpu_no_remote_memory(ptr addrspace(1) %ptr, <2 x half> %value) {
1320 ; COMMON-LABEL: define <2 x half> @test_atomicrmw_fsub_v2f16_global_system__amdgpu_ignore_denormal_mode__amdgpu_no_fine_grained_memory__amdgpu_no_remote_memory(
1321 ; COMMON-SAME: ptr addrspace(1) [[PTR:%.*]], <2 x half> [[VALUE:%.*]]) #[[ATTR0]] {
1322 ; COMMON-NEXT: [[TMP1:%.*]] = load <2 x half>, ptr addrspace(1) [[PTR]], align 4
1323 ; COMMON-NEXT: br label [[ATOMICRMW_START:%.*]]
1324 ; COMMON: atomicrmw.start:
1325 ; COMMON-NEXT: [[LOADED:%.*]] = phi <2 x half> [ [[TMP1]], [[TMP0:%.*]] ], [ [[TMP5:%.*]], [[ATOMICRMW_START]] ]
1326 ; COMMON-NEXT: [[NEW:%.*]] = fsub <2 x half> [[LOADED]], [[VALUE]]
1327 ; COMMON-NEXT: [[TMP2:%.*]] = bitcast <2 x half> [[NEW]] to i32
1328 ; COMMON-NEXT: [[TMP3:%.*]] = bitcast <2 x half> [[LOADED]] to i32
1329 ; COMMON-NEXT: [[TMP4:%.*]] = cmpxchg ptr addrspace(1) [[PTR]], i32 [[TMP3]], i32 [[TMP2]] seq_cst seq_cst, align 4, !amdgpu.no.fine.grained.memory [[META0]], !amdgpu.no.remote.memory [[META0]]
1330 ; COMMON-NEXT: [[SUCCESS:%.*]] = extractvalue { i32, i1 } [[TMP4]], 1
1331 ; COMMON-NEXT: [[NEWLOADED:%.*]] = extractvalue { i32, i1 } [[TMP4]], 0
1332 ; COMMON-NEXT: [[TMP5]] = bitcast i32 [[NEWLOADED]] to <2 x half>
1333 ; COMMON-NEXT: br i1 [[SUCCESS]], label [[ATOMICRMW_END:%.*]], label [[ATOMICRMW_START]]
1334 ; COMMON: atomicrmw.end:
1335 ; COMMON-NEXT: ret <2 x half> [[TMP5]]
1336 ;
1337 %res = atomicrmw fsub ptr addrspace(1) %ptr, <2 x half> %value seq_cst, align 4, !amdgpu.no.fine.grained.memory !0, !amdgpu.no.remote.memory !0, !amdgpu.ignore.denormal.mode !0
1338 ret <2 x half> %res
1339 }
1341 ;---------------------------------------------------------------------
1342 ; atomicrmw fmax
1343 ;---------------------------------------------------------------------
1345 define <2 x half> @test_atomicrmw_fmax_v2f16_global_system(ptr addrspace(1) %ptr, <2 x half> %value) {
1346 ; COMMON-LABEL: define <2 x half> @test_atomicrmw_fmax_v2f16_global_system(
1347 ; COMMON-SAME: ptr addrspace(1) [[PTR:%.*]], <2 x half> [[VALUE:%.*]]) #[[ATTR0]] {
1348 ; COMMON-NEXT: [[TMP1:%.*]] = load <2 x half>, ptr addrspace(1) [[PTR]], align 4
1349 ; COMMON-NEXT: br label [[ATOMICRMW_START:%.*]]
1350 ; COMMON: atomicrmw.start:
1351 ; COMMON-NEXT: [[LOADED:%.*]] = phi <2 x half> [ [[TMP1]], [[TMP0:%.*]] ], [ [[RES:%.*]], [[ATOMICRMW_START]] ]
1352 ; COMMON-NEXT: [[TMP2:%.*]] = call <2 x half> @llvm.maxnum.v2f16(<2 x half> [[LOADED]], <2 x half> [[VALUE]])
1353 ; COMMON-NEXT: [[TMP3:%.*]] = bitcast <2 x half> [[TMP2]] to i32
1354 ; COMMON-NEXT: [[TMP4:%.*]] = bitcast <2 x half> [[LOADED]] to i32
1355 ; COMMON-NEXT: [[TMP5:%.*]] = cmpxchg ptr addrspace(1) [[PTR]], i32 [[TMP4]], i32 [[TMP3]] seq_cst seq_cst, align 4
1356 ; COMMON-NEXT: [[SUCCESS:%.*]] = extractvalue { i32, i1 } [[TMP5]], 1
1357 ; COMMON-NEXT: [[NEWLOADED:%.*]] = extractvalue { i32, i1 } [[TMP5]], 0
1358 ; COMMON-NEXT: [[RES]] = bitcast i32 [[NEWLOADED]] to <2 x half>
1359 ; COMMON-NEXT: br i1 [[SUCCESS]], label [[ATOMICRMW_END:%.*]], label [[ATOMICRMW_START]]
1360 ; COMMON: atomicrmw.end:
1361 ; COMMON-NEXT: ret <2 x half> [[RES]]
1362 ;
1363 %res = atomicrmw fmax ptr addrspace(1) %ptr, <2 x half> %value seq_cst
1364 ret <2 x half> %res
1365 }
1367 define <2 x half> @test_atomicrmw_fmax_v2f16_global_system__amdgpu_no_fine_grained_memory(ptr addrspace(1) %ptr, <2 x half> %value) {
1368 ; COMMON-LABEL: define <2 x half> @test_atomicrmw_fmax_v2f16_global_system__amdgpu_no_fine_grained_memory(
1369 ; COMMON-SAME: ptr addrspace(1) [[PTR:%.*]], <2 x half> [[VALUE:%.*]]) #[[ATTR0]] {
1370 ; COMMON-NEXT: [[TMP1:%.*]] = load <2 x half>, ptr addrspace(1) [[PTR]], align 4
1371 ; COMMON-NEXT: br label [[ATOMICRMW_START:%.*]]
1372 ; COMMON: atomicrmw.start:
1373 ; COMMON-NEXT: [[LOADED:%.*]] = phi <2 x half> [ [[TMP1]], [[TMP0:%.*]] ], [ [[RES:%.*]], [[ATOMICRMW_START]] ]
1374 ; COMMON-NEXT: [[TMP2:%.*]] = call <2 x half> @llvm.maxnum.v2f16(<2 x half> [[LOADED]], <2 x half> [[VALUE]])
1375 ; COMMON-NEXT: [[TMP3:%.*]] = bitcast <2 x half> [[TMP2]] to i32
1376 ; COMMON-NEXT: [[TMP4:%.*]] = bitcast <2 x half> [[LOADED]] to i32
1377 ; COMMON-NEXT: [[TMP5:%.*]] = cmpxchg ptr addrspace(1) [[PTR]], i32 [[TMP4]], i32 [[TMP3]] seq_cst seq_cst, align 4, !amdgpu.no.fine.grained.memory [[META0]]
1378 ; COMMON-NEXT: [[SUCCESS:%.*]] = extractvalue { i32, i1 } [[TMP5]], 1
1379 ; COMMON-NEXT: [[NEWLOADED:%.*]] = extractvalue { i32, i1 } [[TMP5]], 0
1380 ; COMMON-NEXT: [[RES]] = bitcast i32 [[NEWLOADED]] to <2 x half>
1381 ; COMMON-NEXT: br i1 [[SUCCESS]], label [[ATOMICRMW_END:%.*]], label [[ATOMICRMW_START]]
1382 ; COMMON: atomicrmw.end:
1383 ; COMMON-NEXT: ret <2 x half> [[RES]]
1384 ;
1385 %res = atomicrmw fmax ptr addrspace(1) %ptr, <2 x half> %value seq_cst, !amdgpu.no.fine.grained.memory !0
1386 ret <2 x half> %res
1387 }
1389 define <2 x half> @test_atomicrmw_fmax_v2f16_global_system__amdgpu_no_remote_memory(ptr addrspace(1) %ptr, <2 x half> %value) {
1390 ; COMMON-LABEL: define <2 x half> @test_atomicrmw_fmax_v2f16_global_system__amdgpu_no_remote_memory(
1391 ; COMMON-SAME: ptr addrspace(1) [[PTR:%.*]], <2 x half> [[VALUE:%.*]]) #[[ATTR0]] {
1392 ; COMMON-NEXT: [[TMP1:%.*]] = load <2 x half>, ptr addrspace(1) [[PTR]], align 4
1393 ; COMMON-NEXT: br label [[ATOMICRMW_START:%.*]]
1394 ; COMMON: atomicrmw.start:
1395 ; COMMON-NEXT: [[LOADED:%.*]] = phi <2 x half> [ [[TMP1]], [[TMP0:%.*]] ], [ [[RES:%.*]], [[ATOMICRMW_START]] ]
1396 ; COMMON-NEXT: [[TMP2:%.*]] = call <2 x half> @llvm.maxnum.v2f16(<2 x half> [[LOADED]], <2 x half> [[VALUE]])
1397 ; COMMON-NEXT: [[TMP3:%.*]] = bitcast <2 x half> [[TMP2]] to i32
1398 ; COMMON-NEXT: [[TMP4:%.*]] = bitcast <2 x half> [[LOADED]] to i32
1399 ; COMMON-NEXT: [[TMP5:%.*]] = cmpxchg ptr addrspace(1) [[PTR]], i32 [[TMP4]], i32 [[TMP3]] seq_cst seq_cst, align 4, !amdgpu.no.remote.memory [[META0]]
1400 ; COMMON-NEXT: [[SUCCESS:%.*]] = extractvalue { i32, i1 } [[TMP5]], 1
1401 ; COMMON-NEXT: [[NEWLOADED:%.*]] = extractvalue { i32, i1 } [[TMP5]], 0
1402 ; COMMON-NEXT: [[RES]] = bitcast i32 [[NEWLOADED]] to <2 x half>
1403 ; COMMON-NEXT: br i1 [[SUCCESS]], label [[ATOMICRMW_END:%.*]], label [[ATOMICRMW_START]]
1404 ; COMMON: atomicrmw.end:
1405 ; COMMON-NEXT: ret <2 x half> [[RES]]
1406 ;
1407 %res = atomicrmw fmax ptr addrspace(1) %ptr, <2 x half> %value seq_cst, !amdgpu.no.remote.memory !0
1408 ret <2 x half> %res
1409 }
1411 define <2 x half> @test_atomicrmw_fmax_v2f16_global_system__amdgpu_no_fine_grained_memory__amdgpu_no_remote_memory(ptr addrspace(1) %ptr, <2 x half> %value) {
1412 ; COMMON-LABEL: define <2 x half> @test_atomicrmw_fmax_v2f16_global_system__amdgpu_no_fine_grained_memory__amdgpu_no_remote_memory(
1413 ; COMMON-SAME: ptr addrspace(1) [[PTR:%.*]], <2 x half> [[VALUE:%.*]]) #[[ATTR0]] {
1414 ; COMMON-NEXT: [[TMP1:%.*]] = load <2 x half>, ptr addrspace(1) [[PTR]], align 4
1415 ; COMMON-NEXT: br label [[ATOMICRMW_START:%.*]]
1416 ; COMMON: atomicrmw.start:
1417 ; COMMON-NEXT: [[LOADED:%.*]] = phi <2 x half> [ [[TMP1]], [[TMP0:%.*]] ], [ [[RES:%.*]], [[ATOMICRMW_START]] ]
1418 ; COMMON-NEXT: [[TMP2:%.*]] = call <2 x half> @llvm.maxnum.v2f16(<2 x half> [[LOADED]], <2 x half> [[VALUE]])
1419 ; COMMON-NEXT: [[TMP3:%.*]] = bitcast <2 x half> [[TMP2]] to i32
1420 ; COMMON-NEXT: [[TMP4:%.*]] = bitcast <2 x half> [[LOADED]] to i32
1421 ; COMMON-NEXT: [[TMP5:%.*]] = cmpxchg ptr addrspace(1) [[PTR]], i32 [[TMP4]], i32 [[TMP3]] seq_cst seq_cst, align 4, !amdgpu.no.fine.grained.memory [[META0]], !amdgpu.no.remote.memory [[META0]]
1422 ; COMMON-NEXT: [[SUCCESS:%.*]] = extractvalue { i32, i1 } [[TMP5]], 1
1423 ; COMMON-NEXT: [[NEWLOADED:%.*]] = extractvalue { i32, i1 } [[TMP5]], 0
1424 ; COMMON-NEXT: [[RES]] = bitcast i32 [[NEWLOADED]] to <2 x half>
1425 ; COMMON-NEXT: br i1 [[SUCCESS]], label [[ATOMICRMW_END:%.*]], label [[ATOMICRMW_START]]
1426 ; COMMON: atomicrmw.end:
1427 ; COMMON-NEXT: ret <2 x half> [[RES]]
1428 ;
1429 %res = atomicrmw fmax ptr addrspace(1) %ptr, <2 x half> %value seq_cst, !amdgpu.no.fine.grained.memory !0, !amdgpu.no.remote.memory !0
1430 ret <2 x half> %res
1431 }
1433 define <2 x half> @test_atomicrmw_fmax_v2f16_global_system__amdgpu_ignore_denormal_mode(ptr addrspace(1) %ptr, <2 x half> %value) {
1434 ; COMMON-LABEL: define <2 x half> @test_atomicrmw_fmax_v2f16_global_system__amdgpu_ignore_denormal_mode(
1435 ; COMMON-SAME: ptr addrspace(1) [[PTR:%.*]], <2 x half> [[VALUE:%.*]]) #[[ATTR0]] {
1436 ; COMMON-NEXT: [[TMP1:%.*]] = load <2 x half>, ptr addrspace(1) [[PTR]], align 4
1437 ; COMMON-NEXT: br label [[ATOMICRMW_START:%.*]]
1438 ; COMMON: atomicrmw.start:
1439 ; COMMON-NEXT: [[LOADED:%.*]] = phi <2 x half> [ [[TMP1]], [[TMP0:%.*]] ], [ [[TMP6:%.*]], [[ATOMICRMW_START]] ]
1440 ; COMMON-NEXT: [[TMP2:%.*]] = call <2 x half> @llvm.maxnum.v2f16(<2 x half> [[LOADED]], <2 x half> [[VALUE]])
1441 ; COMMON-NEXT: [[TMP3:%.*]] = bitcast <2 x half> [[TMP2]] to i32
1442 ; COMMON-NEXT: [[TMP4:%.*]] = bitcast <2 x half> [[LOADED]] to i32
1443 ; COMMON-NEXT: [[TMP5:%.*]] = cmpxchg ptr addrspace(1) [[PTR]], i32 [[TMP4]], i32 [[TMP3]] seq_cst seq_cst, align 4
1444 ; COMMON-NEXT: [[SUCCESS:%.*]] = extractvalue { i32, i1 } [[TMP5]], 1
1445 ; COMMON-NEXT: [[NEWLOADED:%.*]] = extractvalue { i32, i1 } [[TMP5]], 0
1446 ; COMMON-NEXT: [[TMP6]] = bitcast i32 [[NEWLOADED]] to <2 x half>
1447 ; COMMON-NEXT: br i1 [[SUCCESS]], label [[ATOMICRMW_END:%.*]], label [[ATOMICRMW_START]]
1448 ; COMMON: atomicrmw.end:
1449 ; COMMON-NEXT: ret <2 x half> [[TMP6]]
1450 ;
1451 %res = atomicrmw fmax ptr addrspace(1) %ptr, <2 x half> %value seq_cst, align 4, !amdgpu.ignore.denormal.mode !0
1452 ret <2 x half> %res
1453 }
1455 define <2 x half> @test_atomicrmw_fmax_v2f16_global_system__amdgpu_ignore_denormal_mode__amdgpu_no_fine_grained_memory(ptr addrspace(1) %ptr, <2 x half> %value) {
1456 ; COMMON-LABEL: define <2 x half> @test_atomicrmw_fmax_v2f16_global_system__amdgpu_ignore_denormal_mode__amdgpu_no_fine_grained_memory(
1457 ; COMMON-SAME: ptr addrspace(1) [[PTR:%.*]], <2 x half> [[VALUE:%.*]]) #[[ATTR0]] {
1458 ; COMMON-NEXT: [[TMP1:%.*]] = load <2 x half>, ptr addrspace(1) [[PTR]], align 4
1459 ; COMMON-NEXT: br label [[ATOMICRMW_START:%.*]]
1460 ; COMMON: atomicrmw.start:
1461 ; COMMON-NEXT: [[LOADED:%.*]] = phi <2 x half> [ [[TMP1]], [[TMP0:%.*]] ], [ [[TMP6:%.*]], [[ATOMICRMW_START]] ]
1462 ; COMMON-NEXT: [[TMP2:%.*]] = call <2 x half> @llvm.maxnum.v2f16(<2 x half> [[LOADED]], <2 x half> [[VALUE]])
1463 ; COMMON-NEXT: [[TMP3:%.*]] = bitcast <2 x half> [[TMP2]] to i32
1464 ; COMMON-NEXT: [[TMP4:%.*]] = bitcast <2 x half> [[LOADED]] to i32
1465 ; COMMON-NEXT: [[TMP5:%.*]] = cmpxchg ptr addrspace(1) [[PTR]], i32 [[TMP4]], i32 [[TMP3]] seq_cst seq_cst, align 4, !amdgpu.no.fine.grained.memory [[META0]]
1466 ; COMMON-NEXT: [[SUCCESS:%.*]] = extractvalue { i32, i1 } [[TMP5]], 1
1467 ; COMMON-NEXT: [[NEWLOADED:%.*]] = extractvalue { i32, i1 } [[TMP5]], 0
1468 ; COMMON-NEXT: [[TMP6]] = bitcast i32 [[NEWLOADED]] to <2 x half>
1469 ; COMMON-NEXT: br i1 [[SUCCESS]], label [[ATOMICRMW_END:%.*]], label [[ATOMICRMW_START]]
1470 ; COMMON: atomicrmw.end:
1471 ; COMMON-NEXT: ret <2 x half> [[TMP6]]
1472 ;
1473 %res = atomicrmw fmax ptr addrspace(1) %ptr, <2 x half> %value seq_cst, align 4, !amdgpu.no.fine.grained.memory !0, !amdgpu.ignore.denormal.mode !0
1474 ret <2 x half> %res
1475 }
1477 define <2 x half> @test_atomicrmw_fmax_v2f16_global_system__amdgpu_ignore_denormal_mode__amdgpu_no_remote_memory(ptr addrspace(1) %ptr, <2 x half> %value) {
1478 ; COMMON-LABEL: define <2 x half> @test_atomicrmw_fmax_v2f16_global_system__amdgpu_ignore_denormal_mode__amdgpu_no_remote_memory(
1479 ; COMMON-SAME: ptr addrspace(1) [[PTR:%.*]], <2 x half> [[VALUE:%.*]]) #[[ATTR0]] {
1480 ; COMMON-NEXT: [[TMP1:%.*]] = load <2 x half>, ptr addrspace(1) [[PTR]], align 4
1481 ; COMMON-NEXT: br label [[ATOMICRMW_START:%.*]]
1482 ; COMMON: atomicrmw.start:
1483 ; COMMON-NEXT: [[LOADED:%.*]] = phi <2 x half> [ [[TMP1]], [[TMP0:%.*]] ], [ [[TMP6:%.*]], [[ATOMICRMW_START]] ]
1484 ; COMMON-NEXT: [[TMP2:%.*]] = call <2 x half> @llvm.maxnum.v2f16(<2 x half> [[LOADED]], <2 x half> [[VALUE]])
1485 ; COMMON-NEXT: [[TMP3:%.*]] = bitcast <2 x half> [[TMP2]] to i32
1486 ; COMMON-NEXT: [[TMP4:%.*]] = bitcast <2 x half> [[LOADED]] to i32
1487 ; COMMON-NEXT: [[TMP5:%.*]] = cmpxchg ptr addrspace(1) [[PTR]], i32 [[TMP4]], i32 [[TMP3]] seq_cst seq_cst, align 4, !amdgpu.no.remote.memory [[META0]]
1488 ; COMMON-NEXT: [[SUCCESS:%.*]] = extractvalue { i32, i1 } [[TMP5]], 1
1489 ; COMMON-NEXT: [[NEWLOADED:%.*]] = extractvalue { i32, i1 } [[TMP5]], 0
1490 ; COMMON-NEXT: [[TMP6]] = bitcast i32 [[NEWLOADED]] to <2 x half>
1491 ; COMMON-NEXT: br i1 [[SUCCESS]], label [[ATOMICRMW_END:%.*]], label [[ATOMICRMW_START]]
1492 ; COMMON: atomicrmw.end:
1493 ; COMMON-NEXT: ret <2 x half> [[TMP6]]
1494 ;
1495 %res = atomicrmw fmax ptr addrspace(1) %ptr, <2 x half> %value seq_cst, align 4, !amdgpu.no.remote.memory !0, !amdgpu.ignore.denormal.mode !0
1496 ret <2 x half> %res
1497 }
1499 define <2 x half> @test_atomicrmw_fmax_v2f16_global_system__amdgpu_ignore_denormal_mode__amdgpu_no_fine_grained_memory__amdgpu_no_remote_memory(ptr addrspace(1) %ptr, <2 x half> %value) {
1500 ; COMMON-LABEL: define <2 x half> @test_atomicrmw_fmax_v2f16_global_system__amdgpu_ignore_denormal_mode__amdgpu_no_fine_grained_memory__amdgpu_no_remote_memory(
1501 ; COMMON-SAME: ptr addrspace(1) [[PTR:%.*]], <2 x half> [[VALUE:%.*]]) #[[ATTR0]] {
1502 ; COMMON-NEXT: [[TMP1:%.*]] = load <2 x half>, ptr addrspace(1) [[PTR]], align 4
1503 ; COMMON-NEXT: br label [[ATOMICRMW_START:%.*]]
1504 ; COMMON: atomicrmw.start:
1505 ; COMMON-NEXT: [[LOADED:%.*]] = phi <2 x half> [ [[TMP1]], [[TMP0:%.*]] ], [ [[TMP6:%.*]], [[ATOMICRMW_START]] ]
1506 ; COMMON-NEXT: [[TMP2:%.*]] = call <2 x half> @llvm.maxnum.v2f16(<2 x half> [[LOADED]], <2 x half> [[VALUE]])
1507 ; COMMON-NEXT: [[TMP3:%.*]] = bitcast <2 x half> [[TMP2]] to i32
1508 ; COMMON-NEXT: [[TMP4:%.*]] = bitcast <2 x half> [[LOADED]] to i32
1509 ; COMMON-NEXT: [[TMP5:%.*]] = cmpxchg ptr addrspace(1) [[PTR]], i32 [[TMP4]], i32 [[TMP3]] seq_cst seq_cst, align 4, !amdgpu.no.fine.grained.memory [[META0]], !amdgpu.no.remote.memory [[META0]]
1510 ; COMMON-NEXT: [[SUCCESS:%.*]] = extractvalue { i32, i1 } [[TMP5]], 1
1511 ; COMMON-NEXT: [[NEWLOADED:%.*]] = extractvalue { i32, i1 } [[TMP5]], 0
1512 ; COMMON-NEXT: [[TMP6]] = bitcast i32 [[NEWLOADED]] to <2 x half>
1513 ; COMMON-NEXT: br i1 [[SUCCESS]], label [[ATOMICRMW_END:%.*]], label [[ATOMICRMW_START]]
1514 ; COMMON: atomicrmw.end:
1515 ; COMMON-NEXT: ret <2 x half> [[TMP6]]
1516 ;
1517 %res = atomicrmw fmax ptr addrspace(1) %ptr, <2 x half> %value seq_cst, align 4, !amdgpu.no.fine.grained.memory !0, !amdgpu.no.remote.memory !0, !amdgpu.ignore.denormal.mode !0
1518 ret <2 x half> %res
1519 }
1521 ;---------------------------------------------------------------------
1522 ; atomicrmw fmin
1523 ;---------------------------------------------------------------------
1525 define <2 x half> @test_atomicrmw_fmin_v2f16_global_system(ptr addrspace(1) %ptr, <2 x half> %value) {
1526 ; COMMON-LABEL: define <2 x half> @test_atomicrmw_fmin_v2f16_global_system(
1527 ; COMMON-SAME: ptr addrspace(1) [[PTR:%.*]], <2 x half> [[VALUE:%.*]]) #[[ATTR0]] {
1528 ; COMMON-NEXT: [[TMP1:%.*]] = load <2 x half>, ptr addrspace(1) [[PTR]], align 4
1529 ; COMMON-NEXT: br label [[ATOMICRMW_START:%.*]]
1530 ; COMMON: atomicrmw.start:
1531 ; COMMON-NEXT: [[LOADED:%.*]] = phi <2 x half> [ [[TMP1]], [[TMP0:%.*]] ], [ [[RES:%.*]], [[ATOMICRMW_START]] ]
1532 ; COMMON-NEXT: [[TMP2:%.*]] = call <2 x half> @llvm.minnum.v2f16(<2 x half> [[LOADED]], <2 x half> [[VALUE]])
1533 ; COMMON-NEXT: [[TMP3:%.*]] = bitcast <2 x half> [[TMP2]] to i32
1534 ; COMMON-NEXT: [[TMP4:%.*]] = bitcast <2 x half> [[LOADED]] to i32
1535 ; COMMON-NEXT: [[TMP5:%.*]] = cmpxchg ptr addrspace(1) [[PTR]], i32 [[TMP4]], i32 [[TMP3]] seq_cst seq_cst, align 4
1536 ; COMMON-NEXT: [[SUCCESS:%.*]] = extractvalue { i32, i1 } [[TMP5]], 1
1537 ; COMMON-NEXT: [[NEWLOADED:%.*]] = extractvalue { i32, i1 } [[TMP5]], 0
1538 ; COMMON-NEXT: [[RES]] = bitcast i32 [[NEWLOADED]] to <2 x half>
1539 ; COMMON-NEXT: br i1 [[SUCCESS]], label [[ATOMICRMW_END:%.*]], label [[ATOMICRMW_START]]
1540 ; COMMON: atomicrmw.end:
1541 ; COMMON-NEXT: ret <2 x half> [[RES]]
1542 ;
1543 %res = atomicrmw fmin ptr addrspace(1) %ptr, <2 x half> %value seq_cst
1544 ret <2 x half> %res
1545 }
1547 define <2 x half> @test_atomicrmw_fmin_v2f16_global_system__amdgpu_no_fine_grained_memory(ptr addrspace(1) %ptr, <2 x half> %value) {
1548 ; COMMON-LABEL: define <2 x half> @test_atomicrmw_fmin_v2f16_global_system__amdgpu_no_fine_grained_memory(
1549 ; COMMON-SAME: ptr addrspace(1) [[PTR:%.*]], <2 x half> [[VALUE:%.*]]) #[[ATTR0]] {
1550 ; COMMON-NEXT: [[TMP1:%.*]] = load <2 x half>, ptr addrspace(1) [[PTR]], align 4
1551 ; COMMON-NEXT: br label [[ATOMICRMW_START:%.*]]
1552 ; COMMON: atomicrmw.start:
1553 ; COMMON-NEXT: [[LOADED:%.*]] = phi <2 x half> [ [[TMP1]], [[TMP0:%.*]] ], [ [[RES:%.*]], [[ATOMICRMW_START]] ]
1554 ; COMMON-NEXT: [[TMP2:%.*]] = call <2 x half> @llvm.minnum.v2f16(<2 x half> [[LOADED]], <2 x half> [[VALUE]])
1555 ; COMMON-NEXT: [[TMP3:%.*]] = bitcast <2 x half> [[TMP2]] to i32
1556 ; COMMON-NEXT: [[TMP4:%.*]] = bitcast <2 x half> [[LOADED]] to i32
1557 ; COMMON-NEXT: [[TMP5:%.*]] = cmpxchg ptr addrspace(1) [[PTR]], i32 [[TMP4]], i32 [[TMP3]] seq_cst seq_cst, align 4, !amdgpu.no.fine.grained.memory [[META0]]
1558 ; COMMON-NEXT: [[SUCCESS:%.*]] = extractvalue { i32, i1 } [[TMP5]], 1
1559 ; COMMON-NEXT: [[NEWLOADED:%.*]] = extractvalue { i32, i1 } [[TMP5]], 0
1560 ; COMMON-NEXT: [[RES]] = bitcast i32 [[NEWLOADED]] to <2 x half>
1561 ; COMMON-NEXT: br i1 [[SUCCESS]], label [[ATOMICRMW_END:%.*]], label [[ATOMICRMW_START]]
1562 ; COMMON: atomicrmw.end:
1563 ; COMMON-NEXT: ret <2 x half> [[RES]]
1564 ;
1565 %res = atomicrmw fmin ptr addrspace(1) %ptr, <2 x half> %value seq_cst, !amdgpu.no.fine.grained.memory !0
1566 ret <2 x half> %res
1567 }
1569 define <2 x half> @test_atomicrmw_fmin_v2f16_global_system__amdgpu_no_remote_memory(ptr addrspace(1) %ptr, <2 x half> %value) {
1570 ; COMMON-LABEL: define <2 x half> @test_atomicrmw_fmin_v2f16_global_system__amdgpu_no_remote_memory(
1571 ; COMMON-SAME: ptr addrspace(1) [[PTR:%.*]], <2 x half> [[VALUE:%.*]]) #[[ATTR0]] {
1572 ; COMMON-NEXT: [[TMP1:%.*]] = load <2 x half>, ptr addrspace(1) [[PTR]], align 4
1573 ; COMMON-NEXT: br label [[ATOMICRMW_START:%.*]]
1574 ; COMMON: atomicrmw.start:
1575 ; COMMON-NEXT: [[LOADED:%.*]] = phi <2 x half> [ [[TMP1]], [[TMP0:%.*]] ], [ [[RES:%.*]], [[ATOMICRMW_START]] ]
1576 ; COMMON-NEXT: [[TMP2:%.*]] = call <2 x half> @llvm.minnum.v2f16(<2 x half> [[LOADED]], <2 x half> [[VALUE]])
1577 ; COMMON-NEXT: [[TMP3:%.*]] = bitcast <2 x half> [[TMP2]] to i32
1578 ; COMMON-NEXT: [[TMP4:%.*]] = bitcast <2 x half> [[LOADED]] to i32
1579 ; COMMON-NEXT: [[TMP5:%.*]] = cmpxchg ptr addrspace(1) [[PTR]], i32 [[TMP4]], i32 [[TMP3]] seq_cst seq_cst, align 4, !amdgpu.no.remote.memory [[META0]]
1580 ; COMMON-NEXT: [[SUCCESS:%.*]] = extractvalue { i32, i1 } [[TMP5]], 1
1581 ; COMMON-NEXT: [[NEWLOADED:%.*]] = extractvalue { i32, i1 } [[TMP5]], 0
1582 ; COMMON-NEXT: [[RES]] = bitcast i32 [[NEWLOADED]] to <2 x half>
1583 ; COMMON-NEXT: br i1 [[SUCCESS]], label [[ATOMICRMW_END:%.*]], label [[ATOMICRMW_START]]
1584 ; COMMON: atomicrmw.end:
1585 ; COMMON-NEXT: ret <2 x half> [[RES]]
1586 ;
1587 %res = atomicrmw fmin ptr addrspace(1) %ptr, <2 x half> %value seq_cst, !amdgpu.no.remote.memory !0
1588 ret <2 x half> %res
1589 }
1591 define <2 x half> @test_atomicrmw_fmin_v2f16_global_system__amdgpu_no_fine_grained_memory__amdgpu_no_remote_memory(ptr addrspace(1) %ptr, <2 x half> %value) {
1592 ; COMMON-LABEL: define <2 x half> @test_atomicrmw_fmin_v2f16_global_system__amdgpu_no_fine_grained_memory__amdgpu_no_remote_memory(
1593 ; COMMON-SAME: ptr addrspace(1) [[PTR:%.*]], <2 x half> [[VALUE:%.*]]) #[[ATTR0]] {
1594 ; COMMON-NEXT: [[TMP1:%.*]] = load <2 x half>, ptr addrspace(1) [[PTR]], align 4
1595 ; COMMON-NEXT: br label [[ATOMICRMW_START:%.*]]
1596 ; COMMON: atomicrmw.start:
1597 ; COMMON-NEXT: [[LOADED:%.*]] = phi <2 x half> [ [[TMP1]], [[TMP0:%.*]] ], [ [[RES:%.*]], [[ATOMICRMW_START]] ]
1598 ; COMMON-NEXT: [[TMP2:%.*]] = call <2 x half> @llvm.minnum.v2f16(<2 x half> [[LOADED]], <2 x half> [[VALUE]])
1599 ; COMMON-NEXT: [[TMP3:%.*]] = bitcast <2 x half> [[TMP2]] to i32
1600 ; COMMON-NEXT: [[TMP4:%.*]] = bitcast <2 x half> [[LOADED]] to i32
1601 ; COMMON-NEXT: [[TMP5:%.*]] = cmpxchg ptr addrspace(1) [[PTR]], i32 [[TMP4]], i32 [[TMP3]] seq_cst seq_cst, align 4, !amdgpu.no.fine.grained.memory [[META0]], !amdgpu.no.remote.memory [[META0]]
1602 ; COMMON-NEXT: [[SUCCESS:%.*]] = extractvalue { i32, i1 } [[TMP5]], 1
1603 ; COMMON-NEXT: [[NEWLOADED:%.*]] = extractvalue { i32, i1 } [[TMP5]], 0
1604 ; COMMON-NEXT: [[RES]] = bitcast i32 [[NEWLOADED]] to <2 x half>
1605 ; COMMON-NEXT: br i1 [[SUCCESS]], label [[ATOMICRMW_END:%.*]], label [[ATOMICRMW_START]]
1606 ; COMMON: atomicrmw.end:
1607 ; COMMON-NEXT: ret <2 x half> [[RES]]
1608 ;
1609 %res = atomicrmw fmin ptr addrspace(1) %ptr, <2 x half> %value seq_cst, !amdgpu.no.fine.grained.memory !0, !amdgpu.no.remote.memory !0
1610 ret <2 x half> %res
1611 }
1613 define <2 x half> @test_atomicrmw_fmin_v2f16_global_system__amdgpu_ignore_denormal_mode(ptr addrspace(1) %ptr, <2 x half> %value) {
1614 ; COMMON-LABEL: define <2 x half> @test_atomicrmw_fmin_v2f16_global_system__amdgpu_ignore_denormal_mode(
1615 ; COMMON-SAME: ptr addrspace(1) [[PTR:%.*]], <2 x half> [[VALUE:%.*]]) #[[ATTR0]] {
1616 ; COMMON-NEXT: [[TMP1:%.*]] = load <2 x half>, ptr addrspace(1) [[PTR]], align 4
1617 ; COMMON-NEXT: br label [[ATOMICRMW_START:%.*]]
1618 ; COMMON: atomicrmw.start:
1619 ; COMMON-NEXT: [[LOADED:%.*]] = phi <2 x half> [ [[TMP1]], [[TMP0:%.*]] ], [ [[TMP6:%.*]], [[ATOMICRMW_START]] ]
1620 ; COMMON-NEXT: [[TMP2:%.*]] = call <2 x half> @llvm.minnum.v2f16(<2 x half> [[LOADED]], <2 x half> [[VALUE]])
1621 ; COMMON-NEXT: [[TMP3:%.*]] = bitcast <2 x half> [[TMP2]] to i32
1622 ; COMMON-NEXT: [[TMP4:%.*]] = bitcast <2 x half> [[LOADED]] to i32
1623 ; COMMON-NEXT: [[TMP5:%.*]] = cmpxchg ptr addrspace(1) [[PTR]], i32 [[TMP4]], i32 [[TMP3]] seq_cst seq_cst, align 4
1624 ; COMMON-NEXT: [[SUCCESS:%.*]] = extractvalue { i32, i1 } [[TMP5]], 1
1625 ; COMMON-NEXT: [[NEWLOADED:%.*]] = extractvalue { i32, i1 } [[TMP5]], 0
1626 ; COMMON-NEXT: [[TMP6]] = bitcast i32 [[NEWLOADED]] to <2 x half>
1627 ; COMMON-NEXT: br i1 [[SUCCESS]], label [[ATOMICRMW_END:%.*]], label [[ATOMICRMW_START]]
1628 ; COMMON: atomicrmw.end:
1629 ; COMMON-NEXT: ret <2 x half> [[TMP6]]
1630 ;
1631 %res = atomicrmw fmin ptr addrspace(1) %ptr, <2 x half> %value seq_cst, align 4, !amdgpu.ignore.denormal.mode !0
1632 ret <2 x half> %res
1633 }
1635 define <2 x half> @test_atomicrmw_fmin_v2f16_global_system__amdgpu_ignore_denormal_mode__amdgpu_no_fine_grained_memory(ptr addrspace(1) %ptr, <2 x half> %value) {
1636 ; COMMON-LABEL: define <2 x half> @test_atomicrmw_fmin_v2f16_global_system__amdgpu_ignore_denormal_mode__amdgpu_no_fine_grained_memory(
1637 ; COMMON-SAME: ptr addrspace(1) [[PTR:%.*]], <2 x half> [[VALUE:%.*]]) #[[ATTR0]] {
1638 ; COMMON-NEXT: [[TMP1:%.*]] = load <2 x half>, ptr addrspace(1) [[PTR]], align 4
1639 ; COMMON-NEXT: br label [[ATOMICRMW_START:%.*]]
1640 ; COMMON: atomicrmw.start:
1641 ; COMMON-NEXT: [[LOADED:%.*]] = phi <2 x half> [ [[TMP1]], [[TMP0:%.*]] ], [ [[TMP6:%.*]], [[ATOMICRMW_START]] ]
1642 ; COMMON-NEXT: [[TMP2:%.*]] = call <2 x half> @llvm.minnum.v2f16(<2 x half> [[LOADED]], <2 x half> [[VALUE]])
1643 ; COMMON-NEXT: [[TMP3:%.*]] = bitcast <2 x half> [[TMP2]] to i32
1644 ; COMMON-NEXT: [[TMP4:%.*]] = bitcast <2 x half> [[LOADED]] to i32
1645 ; COMMON-NEXT: [[TMP5:%.*]] = cmpxchg ptr addrspace(1) [[PTR]], i32 [[TMP4]], i32 [[TMP3]] seq_cst seq_cst, align 4, !amdgpu.no.fine.grained.memory [[META0]]
1646 ; COMMON-NEXT: [[SUCCESS:%.*]] = extractvalue { i32, i1 } [[TMP5]], 1
1647 ; COMMON-NEXT: [[NEWLOADED:%.*]] = extractvalue { i32, i1 } [[TMP5]], 0
1648 ; COMMON-NEXT: [[TMP6]] = bitcast i32 [[NEWLOADED]] to <2 x half>
1649 ; COMMON-NEXT: br i1 [[SUCCESS]], label [[ATOMICRMW_END:%.*]], label [[ATOMICRMW_START]]
1650 ; COMMON: atomicrmw.end:
1651 ; COMMON-NEXT: ret <2 x half> [[TMP6]]
1652 ;
1653 %res = atomicrmw fmin ptr addrspace(1) %ptr, <2 x half> %value seq_cst, align 4, !amdgpu.no.fine.grained.memory !0, !amdgpu.ignore.denormal.mode !0
1654 ret <2 x half> %res
1655 }
1657 define <2 x half> @test_atomicrmw_fmin_v2f16_global_system__amdgpu_ignore_denormal_mode__amdgpu_no_remote_memory(ptr addrspace(1) %ptr, <2 x half> %value) {
1658 ; COMMON-LABEL: define <2 x half> @test_atomicrmw_fmin_v2f16_global_system__amdgpu_ignore_denormal_mode__amdgpu_no_remote_memory(
1659 ; COMMON-SAME: ptr addrspace(1) [[PTR:%.*]], <2 x half> [[VALUE:%.*]]) #[[ATTR0]] {
1660 ; COMMON-NEXT: [[TMP1:%.*]] = load <2 x half>, ptr addrspace(1) [[PTR]], align 4
1661 ; COMMON-NEXT: br label [[ATOMICRMW_START:%.*]]
1662 ; COMMON: atomicrmw.start:
1663 ; COMMON-NEXT: [[LOADED:%.*]] = phi <2 x half> [ [[TMP1]], [[TMP0:%.*]] ], [ [[TMP6:%.*]], [[ATOMICRMW_START]] ]
1664 ; COMMON-NEXT: [[TMP2:%.*]] = call <2 x half> @llvm.minnum.v2f16(<2 x half> [[LOADED]], <2 x half> [[VALUE]])
1665 ; COMMON-NEXT: [[TMP3:%.*]] = bitcast <2 x half> [[TMP2]] to i32
1666 ; COMMON-NEXT: [[TMP4:%.*]] = bitcast <2 x half> [[LOADED]] to i32
1667 ; COMMON-NEXT: [[TMP5:%.*]] = cmpxchg ptr addrspace(1) [[PTR]], i32 [[TMP4]], i32 [[TMP3]] seq_cst seq_cst, align 4, !amdgpu.no.remote.memory [[META0]]
1668 ; COMMON-NEXT: [[SUCCESS:%.*]] = extractvalue { i32, i1 } [[TMP5]], 1
1669 ; COMMON-NEXT: [[NEWLOADED:%.*]] = extractvalue { i32, i1 } [[TMP5]], 0
1670 ; COMMON-NEXT: [[TMP6]] = bitcast i32 [[NEWLOADED]] to <2 x half>
1671 ; COMMON-NEXT: br i1 [[SUCCESS]], label [[ATOMICRMW_END:%.*]], label [[ATOMICRMW_START]]
1672 ; COMMON: atomicrmw.end:
1673 ; COMMON-NEXT: ret <2 x half> [[TMP6]]
1674 ;
1675 %res = atomicrmw fmin ptr addrspace(1) %ptr, <2 x half> %value seq_cst, align 4, !amdgpu.no.remote.memory !0, !amdgpu.ignore.denormal.mode !0
1676 ret <2 x half> %res
1677 }
1679 define <2 x half> @test_atomicrmw_fmin_v2f16_global_system__amdgpu_ignore_denormal_mode__amdgpu_no_fine_grained_memory__amdgpu_no_remote_memory(ptr addrspace(1) %ptr, <2 x half> %value) {
1680 ; COMMON-LABEL: define <2 x half> @test_atomicrmw_fmin_v2f16_global_system__amdgpu_ignore_denormal_mode__amdgpu_no_fine_grained_memory__amdgpu_no_remote_memory(
1681 ; COMMON-SAME: ptr addrspace(1) [[PTR:%.*]], <2 x half> [[VALUE:%.*]]) #[[ATTR0]] {
1682 ; COMMON-NEXT: [[TMP1:%.*]] = load <2 x half>, ptr addrspace(1) [[PTR]], align 4
1683 ; COMMON-NEXT: br label [[ATOMICRMW_START:%.*]]
1684 ; COMMON: atomicrmw.start:
1685 ; COMMON-NEXT: [[LOADED:%.*]] = phi <2 x half> [ [[TMP1]], [[TMP0:%.*]] ], [ [[TMP6:%.*]], [[ATOMICRMW_START]] ]
1686 ; COMMON-NEXT: [[TMP2:%.*]] = call <2 x half> @llvm.minnum.v2f16(<2 x half> [[LOADED]], <2 x half> [[VALUE]])
1687 ; COMMON-NEXT: [[TMP3:%.*]] = bitcast <2 x half> [[TMP2]] to i32
1688 ; COMMON-NEXT: [[TMP4:%.*]] = bitcast <2 x half> [[LOADED]] to i32
1689 ; COMMON-NEXT: [[TMP5:%.*]] = cmpxchg ptr addrspace(1) [[PTR]], i32 [[TMP4]], i32 [[TMP3]] seq_cst seq_cst, align 4, !amdgpu.no.fine.grained.memory [[META0]], !amdgpu.no.remote.memory [[META0]]
1690 ; COMMON-NEXT: [[SUCCESS:%.*]] = extractvalue { i32, i1 } [[TMP5]], 1
1691 ; COMMON-NEXT: [[NEWLOADED:%.*]] = extractvalue { i32, i1 } [[TMP5]], 0
1692 ; COMMON-NEXT: [[TMP6]] = bitcast i32 [[NEWLOADED]] to <2 x half>
1693 ; COMMON-NEXT: br i1 [[SUCCESS]], label [[ATOMICRMW_END:%.*]], label [[ATOMICRMW_START]]
1694 ; COMMON: atomicrmw.end:
1695 ; COMMON-NEXT: ret <2 x half> [[TMP6]]
1696 ;
1697 %res = atomicrmw fmin ptr addrspace(1) %ptr, <2 x half> %value seq_cst, align 4, !amdgpu.no.fine.grained.memory !0, !amdgpu.no.remote.memory !0, !amdgpu.ignore.denormal.mode !0
1698 ret <2 x half> %res
1699 }
1701 attributes #0 = { "denormal-fp-mode"="preserve-sign,preserve-sign" }
1702 attributes #1 = { "denormal-fp-mode"="dynamic,dynamic" }
1704 !0 = !{}
1705 ;.
1706 ; GFX803: [[META0]] = !{}
1707 ;.
1708 ; GFX906: [[META0]] = !{}
1709 ;.
1710 ; GFX908: [[META0]] = !{}
1711 ;.
1712 ; GFX90A: [[META0]] = !{}
1713 ;.
1714 ; GFX940: [[META0]] = !{}
1715 ;.
1716 ; GFX10: [[META0]] = !{}
1717 ;.
1718 ; GFX11: [[META0]] = !{}
1719 ;.
1720 ; GFX12: [[META0]] = !{}
1721 ;.