1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197
198
199
200
201
202
203
204
205
206
207
208
209
210
211
212
213
214
215
216
217
218
219
220
221
222
223
224
225
226
227
228
229
230
231
232
233
234
235
236
237
238
239
240
241
242
243
244
245
246
247
248
249
250
251
252
253
254
255
256
257
258
259
260
261
262
263
264
265
266
267
|
; NOTE: Assertions have been autogenerated by utils/update_llc_test_checks.py
; RUN: llc -aarch64-enable-mgather-combine=0 -enable-misched=false < %s | FileCheck %s
; RUN: llc -aarch64-enable-mgather-combine=1 -enable-misched=false < %s | FileCheck %s
target triple = "aarch64-linux-gnu"
; Test for multiple uses of the mgather where the s/zext should not be combined
define <vscale x 2 x i64> @masked_sgather_sext(ptr %base, <vscale x 2 x i64> %offsets, <vscale x 2 x i1> %mask, <vscale x 2 x i8> %vals) #0 {
; CHECK-LABEL: masked_sgather_sext:
; CHECK: // %bb.0:
; CHECK-NEXT: ld1b { z0.d }, p0/z, [x0, z0.d]
; CHECK-NEXT: ptrue p0.d
; CHECK-NEXT: movprfx z2, z0
; CHECK-NEXT: sxtb z2.d, p0/m, z0.d
; CHECK-NEXT: add z0.d, z0.d, z1.d
; CHECK-NEXT: sxtb z0.d, p0/m, z0.d
; CHECK-NEXT: mul z0.d, p0/m, z0.d, z2.d
; CHECK-NEXT: ret
%ptrs = getelementptr i8, ptr %base, <vscale x 2 x i64> %offsets
%data = call <vscale x 2 x i8> @llvm.masked.gather.nxv2i8(<vscale x 2 x ptr> %ptrs, i32 1, <vscale x 2 x i1> %mask, <vscale x 2 x i8> undef)
%data.sext = sext <vscale x 2 x i8> %data to <vscale x 2 x i64>
%add = add <vscale x 2 x i8> %data, %vals
%add.sext = sext <vscale x 2 x i8> %add to <vscale x 2 x i64>
%mul = mul <vscale x 2 x i64> %data.sext, %add.sext
ret <vscale x 2 x i64> %mul
}
define <vscale x 2 x i64> @masked_sgather_zext(ptr %base, <vscale x 2 x i64> %offsets, <vscale x 2 x i1> %mask, <vscale x 2 x i8> %vals) #0 {
; CHECK-LABEL: masked_sgather_zext:
; CHECK: // %bb.0:
; CHECK-NEXT: ld1b { z0.d }, p0/z, [x0, z0.d]
; CHECK-NEXT: ptrue p0.d
; CHECK-NEXT: add z1.d, z0.d, z1.d
; CHECK-NEXT: and z0.d, z0.d, #0xff
; CHECK-NEXT: and z1.d, z1.d, #0xff
; CHECK-NEXT: mul z0.d, p0/m, z0.d, z1.d
; CHECK-NEXT: ret
%ptrs = getelementptr i8, ptr %base, <vscale x 2 x i64> %offsets
%data = call <vscale x 2 x i8> @llvm.masked.gather.nxv2i8(<vscale x 2 x ptr> %ptrs, i32 1, <vscale x 2 x i1> %mask, <vscale x 2 x i8> undef)
%data.zext = zext <vscale x 2 x i8> %data to <vscale x 2 x i64>
%add = add <vscale x 2 x i8> %data, %vals
%add.zext = zext <vscale x 2 x i8> %add to <vscale x 2 x i64>
%mul = mul <vscale x 2 x i64> %data.zext, %add.zext
ret <vscale x 2 x i64> %mul
}
; Tests that exercise various type legalisation scenarios for ISD::MGATHER.
; Code generate load of an illegal datatype via promotion.
define <vscale x 2 x i8> @masked_gather_nxv2i8(<vscale x 2 x ptr> %ptrs, <vscale x 2 x i1> %mask) #0 {
; CHECK-LABEL: masked_gather_nxv2i8:
; CHECK: // %bb.0:
; CHECK-NEXT: ld1b { z0.d }, p0/z, [z0.d]
; CHECK-NEXT: ret
%data = call <vscale x 2 x i8> @llvm.masked.gather.nxv2i8(<vscale x 2 x ptr> %ptrs, i32 1, <vscale x 2 x i1> %mask, <vscale x 2 x i8> undef)
ret <vscale x 2 x i8> %data
}
; Code generate load of an illegal datatype via promotion.
define <vscale x 2 x i16> @masked_gather_nxv2i16(<vscale x 2 x ptr> %ptrs, <vscale x 2 x i1> %mask) #0 {
; CHECK-LABEL: masked_gather_nxv2i16:
; CHECK: // %bb.0:
; CHECK-NEXT: ld1h { z0.d }, p0/z, [z0.d]
; CHECK-NEXT: ret
%data = call <vscale x 2 x i16> @llvm.masked.gather.nxv2i16(<vscale x 2 x ptr> %ptrs, i32 2, <vscale x 2 x i1> %mask, <vscale x 2 x i16> undef)
ret <vscale x 2 x i16> %data
}
; Code generate load of an illegal datatype via promotion.
define <vscale x 2 x i32> @masked_gather_nxv2i32(<vscale x 2 x ptr> %ptrs, <vscale x 2 x i1> %mask) #0 {
; CHECK-LABEL: masked_gather_nxv2i32:
; CHECK: // %bb.0:
; CHECK-NEXT: ld1w { z0.d }, p0/z, [z0.d]
; CHECK-NEXT: ret
%data = call <vscale x 2 x i32> @llvm.masked.gather.nxv2i32(<vscale x 2 x ptr> %ptrs, i32 4, <vscale x 2 x i1> %mask, <vscale x 2 x i32> undef)
ret <vscale x 2 x i32> %data
}
define <vscale x 4 x half> @masked_gather_nxv4f16(<vscale x 4 x ptr> %ptrs, <vscale x 4 x i1> %mask) #0 {
; CHECK-LABEL: masked_gather_nxv4f16:
; CHECK: // %bb.0:
; CHECK-NEXT: punpkhi p1.h, p0.b
; CHECK-NEXT: punpklo p0.h, p0.b
; CHECK-NEXT: ld1h { z1.d }, p1/z, [z1.d]
; CHECK-NEXT: ld1h { z0.d }, p0/z, [z0.d]
; CHECK-NEXT: uzp1 z0.s, z0.s, z1.s
; CHECK-NEXT: ret
%data = call <vscale x 4 x half> @llvm.masked.gather.nxv4f16(<vscale x 4 x ptr> %ptrs, i32 0, <vscale x 4 x i1> %mask, <vscale x 4 x half> undef)
ret <vscale x 4 x half> %data
}
define <vscale x 2 x float> @masked_gather_nxv2f32(ptr %base, <vscale x 2 x i16> %indices, <vscale x 2 x i1> %mask) #0 {
; CHECK-LABEL: masked_gather_nxv2f32:
; CHECK: // %bb.0:
; CHECK-NEXT: ptrue p1.d
; CHECK-NEXT: sxth z0.d, p1/m, z0.d
; CHECK-NEXT: ld1w { z0.d }, p0/z, [x0, z0.d, lsl #2]
; CHECK-NEXT: ret
%ptrs = getelementptr float, ptr %base, <vscale x 2 x i16> %indices
%data = call <vscale x 2 x float> @llvm.masked.gather.nxv2f32(<vscale x 2 x ptr> %ptrs, i32 1, <vscale x 2 x i1> %mask, <vscale x 2 x float> undef)
ret <vscale x 2 x float> %data
}
define <vscale x 8 x half> @masked_gather_nxv8f16(<vscale x 8 x ptr> %ptrs, <vscale x 8 x i1> %mask) #0 {
; CHECK-LABEL: masked_gather_nxv8f16:
; CHECK: // %bb.0:
; CHECK-NEXT: punpkhi p1.h, p0.b
; CHECK-NEXT: punpklo p0.h, p0.b
; CHECK-NEXT: punpkhi p2.h, p1.b
; CHECK-NEXT: punpklo p1.h, p1.b
; CHECK-NEXT: ld1h { z3.d }, p2/z, [z3.d]
; CHECK-NEXT: ld1h { z2.d }, p1/z, [z2.d]
; CHECK-NEXT: punpkhi p1.h, p0.b
; CHECK-NEXT: punpklo p0.h, p0.b
; CHECK-NEXT: ld1h { z1.d }, p1/z, [z1.d]
; CHECK-NEXT: ld1h { z0.d }, p0/z, [z0.d]
; CHECK-NEXT: uzp1 z2.s, z2.s, z3.s
; CHECK-NEXT: uzp1 z0.s, z0.s, z1.s
; CHECK-NEXT: uzp1 z0.h, z0.h, z2.h
; CHECK-NEXT: ret
%data = call <vscale x 8 x half> @llvm.masked.gather.nxv8f16(<vscale x 8 x ptr> %ptrs, i32 2, <vscale x 8 x i1> %mask, <vscale x 8 x half> undef)
ret <vscale x 8 x half> %data
}
define <vscale x 8 x bfloat> @masked_gather_nxv8bf16(ptr %base, <vscale x 8 x i16> %indices, <vscale x 8 x i1> %mask) #0 {
; CHECK-LABEL: masked_gather_nxv8bf16:
; CHECK: // %bb.0:
; CHECK-NEXT: sunpkhi z1.s, z0.h
; CHECK-NEXT: punpkhi p1.h, p0.b
; CHECK-NEXT: sunpklo z0.s, z0.h
; CHECK-NEXT: punpklo p0.h, p0.b
; CHECK-NEXT: ld1h { z1.s }, p1/z, [x0, z1.s, sxtw #1]
; CHECK-NEXT: ld1h { z0.s }, p0/z, [x0, z0.s, sxtw #1]
; CHECK-NEXT: uzp1 z0.h, z0.h, z1.h
; CHECK-NEXT: ret
%ptrs = getelementptr bfloat, ptr %base, <vscale x 8 x i16> %indices
%data = call <vscale x 8 x bfloat> @llvm.masked.gather.nxv8bf16(<vscale x 8 x ptr> %ptrs, i32 1, <vscale x 8 x i1> %mask, <vscale x 8 x bfloat> undef)
ret <vscale x 8 x bfloat> %data
}
define <vscale x 4 x double> @masked_gather_nxv4f64(ptr %base, <vscale x 4 x i16> %indices, <vscale x 4 x i1> %mask) #0 {
; CHECK-LABEL: masked_gather_nxv4f64:
; CHECK: // %bb.0:
; CHECK-NEXT: ptrue p1.s
; CHECK-NEXT: movprfx z1, z0
; CHECK-NEXT: sxth z1.s, p1/m, z0.s
; CHECK-NEXT: sunpklo z0.d, z1.s
; CHECK-NEXT: punpklo p1.h, p0.b
; CHECK-NEXT: sunpkhi z1.d, z1.s
; CHECK-NEXT: punpkhi p0.h, p0.b
; CHECK-NEXT: ld1d { z0.d }, p1/z, [x0, z0.d, lsl #3]
; CHECK-NEXT: ld1d { z1.d }, p0/z, [x0, z1.d, lsl #3]
; CHECK-NEXT: ret
%ptrs = getelementptr double, ptr %base, <vscale x 4 x i16> %indices
%data = call <vscale x 4 x double> @llvm.masked.gather.nxv4f64(<vscale x 4 x ptr> %ptrs, i32 1, <vscale x 4 x i1> %mask, <vscale x 4 x double> undef)
ret <vscale x 4 x double> %data
}
define <vscale x 8 x float> @masked_gather_nxv8f32(ptr %base, <vscale x 8 x i32> %offsets, <vscale x 8 x i1> %mask) #0 {
; CHECK-LABEL: masked_gather_nxv8f32:
; CHECK: // %bb.0:
; CHECK-NEXT: punpklo p1.h, p0.b
; CHECK-NEXT: punpkhi p0.h, p0.b
; CHECK-NEXT: ld1w { z0.s }, p1/z, [x0, z0.s, uxtw #2]
; CHECK-NEXT: ld1w { z1.s }, p0/z, [x0, z1.s, uxtw #2]
; CHECK-NEXT: ret
%offsets.zext = zext <vscale x 8 x i32> %offsets to <vscale x 8 x i64>
%ptrs = getelementptr float, ptr %base, <vscale x 8 x i64> %offsets.zext
%vals = call <vscale x 8 x float> @llvm.masked.gather.nxv8f32(<vscale x 8 x ptr> %ptrs, i32 4, <vscale x 8 x i1> %mask, <vscale x 8 x float> undef)
ret <vscale x 8 x float> %vals
}
; Code generate the worst case scenario when all vector types are legal.
define <vscale x 16 x i8> @masked_gather_nxv16i8(ptr %base, <vscale x 16 x i8> %indices, <vscale x 16 x i1> %mask) #0 {
; CHECK-LABEL: masked_gather_nxv16i8:
; CHECK: // %bb.0:
; CHECK-NEXT: sunpkhi z1.h, z0.b
; CHECK-NEXT: punpkhi p1.h, p0.b
; CHECK-NEXT: sunpkhi z2.s, z1.h
; CHECK-NEXT: punpkhi p2.h, p1.b
; CHECK-NEXT: sunpklo z1.s, z1.h
; CHECK-NEXT: punpklo p1.h, p1.b
; CHECK-NEXT: ld1b { z2.s }, p2/z, [x0, z2.s, sxtw]
; CHECK-NEXT: ld1b { z1.s }, p1/z, [x0, z1.s, sxtw]
; CHECK-NEXT: sunpklo z0.h, z0.b
; CHECK-NEXT: punpklo p0.h, p0.b
; CHECK-NEXT: punpkhi p1.h, p0.b
; CHECK-NEXT: punpklo p0.h, p0.b
; CHECK-NEXT: uzp1 z1.h, z1.h, z2.h
; CHECK-NEXT: sunpkhi z2.s, z0.h
; CHECK-NEXT: sunpklo z0.s, z0.h
; CHECK-NEXT: ld1b { z2.s }, p1/z, [x0, z2.s, sxtw]
; CHECK-NEXT: ld1b { z0.s }, p0/z, [x0, z0.s, sxtw]
; CHECK-NEXT: uzp1 z0.h, z0.h, z2.h
; CHECK-NEXT: uzp1 z0.b, z0.b, z1.b
; CHECK-NEXT: ret
%ptrs = getelementptr i8, ptr %base, <vscale x 16 x i8> %indices
%data = call <vscale x 16 x i8> @llvm.masked.gather.nxv16i8(<vscale x 16 x ptr> %ptrs, i32 1, <vscale x 16 x i1> %mask, <vscale x 16 x i8> undef)
ret <vscale x 16 x i8> %data
}
; Code generate the worst case scenario when all vector types are illegal.
define <vscale x 32 x i32> @masked_gather_nxv32i32(ptr %base, <vscale x 32 x i32> %indices, <vscale x 32 x i1> %mask) #0 {
; CHECK-LABEL: masked_gather_nxv32i32:
; CHECK: // %bb.0:
; CHECK-NEXT: punpklo p2.h, p0.b
; CHECK-NEXT: punpkhi p0.h, p0.b
; CHECK-NEXT: punpklo p3.h, p2.b
; CHECK-NEXT: punpkhi p2.h, p2.b
; CHECK-NEXT: ld1w { z0.s }, p3/z, [x0, z0.s, sxtw #2]
; CHECK-NEXT: ld1w { z1.s }, p2/z, [x0, z1.s, sxtw #2]
; CHECK-NEXT: punpklo p2.h, p0.b
; CHECK-NEXT: punpkhi p0.h, p0.b
; CHECK-NEXT: ld1w { z2.s }, p2/z, [x0, z2.s, sxtw #2]
; CHECK-NEXT: ld1w { z3.s }, p0/z, [x0, z3.s, sxtw #2]
; CHECK-NEXT: punpklo p0.h, p1.b
; CHECK-NEXT: punpklo p2.h, p0.b
; CHECK-NEXT: punpkhi p0.h, p0.b
; CHECK-NEXT: ld1w { z4.s }, p2/z, [x0, z4.s, sxtw #2]
; CHECK-NEXT: ld1w { z5.s }, p0/z, [x0, z5.s, sxtw #2]
; CHECK-NEXT: punpkhi p0.h, p1.b
; CHECK-NEXT: punpklo p1.h, p0.b
; CHECK-NEXT: punpkhi p0.h, p0.b
; CHECK-NEXT: ld1w { z6.s }, p1/z, [x0, z6.s, sxtw #2]
; CHECK-NEXT: ld1w { z7.s }, p0/z, [x0, z7.s, sxtw #2]
; CHECK-NEXT: ret
%ptrs = getelementptr i32, ptr %base, <vscale x 32 x i32> %indices
%data = call <vscale x 32 x i32> @llvm.masked.gather.nxv32i32(<vscale x 32 x ptr> %ptrs, i32 4, <vscale x 32 x i1> %mask, <vscale x 32 x i32> undef)
ret <vscale x 32 x i32> %data
}
; TODO: Currently, the sign extend gets applied to the values after a 'uzp1' of two
; registers, so it doesn't get folded away. Same for any other vector-of-pointers
; style gathers which don't fit in an <vscale x 2 x ptr> single register. Better folding
; is required before we can check those off.
define <vscale x 4 x i32> @masked_sgather_nxv4i8(<vscale x 4 x ptr> %ptrs, <vscale x 4 x i1> %mask) #0 {
; CHECK-LABEL: masked_sgather_nxv4i8:
; CHECK: // %bb.0:
; CHECK-NEXT: punpkhi p1.h, p0.b
; CHECK-NEXT: punpklo p0.h, p0.b
; CHECK-NEXT: ld1b { z1.d }, p1/z, [z1.d]
; CHECK-NEXT: ld1b { z0.d }, p0/z, [z0.d]
; CHECK-NEXT: ptrue p0.s
; CHECK-NEXT: uzp1 z0.s, z0.s, z1.s
; CHECK-NEXT: sxtb z0.s, p0/m, z0.s
; CHECK-NEXT: ret
%vals = call <vscale x 4 x i8> @llvm.masked.gather.nxv4i8(<vscale x 4 x ptr> %ptrs, i32 1, <vscale x 4 x i1> %mask, <vscale x 4 x i8> undef)
%svals = sext <vscale x 4 x i8> %vals to <vscale x 4 x i32>
ret <vscale x 4 x i32> %svals
}
attributes #0 = { nounwind "target-features"="+sve,+bf16" }
declare <vscale x 2 x i8> @llvm.masked.gather.nxv2i8(<vscale x 2 x ptr>, i32, <vscale x 2 x i1>, <vscale x 2 x i8>)
declare <vscale x 2 x i16> @llvm.masked.gather.nxv2i16(<vscale x 2 x ptr>, i32, <vscale x 2 x i1>, <vscale x 2 x i16>)
declare <vscale x 2 x i32> @llvm.masked.gather.nxv2i32(<vscale x 2 x ptr>, i32, <vscale x 2 x i1>, <vscale x 2 x i32>)
declare <vscale x 4 x i8> @llvm.masked.gather.nxv4i8(<vscale x 4 x ptr>, i32, <vscale x 4 x i1>, <vscale x 4 x i8>)
declare <vscale x 16 x i8> @llvm.masked.gather.nxv16i8(<vscale x 16 x ptr>, i32, <vscale x 16 x i1>, <vscale x 16 x i8>)
declare <vscale x 32 x i32> @llvm.masked.gather.nxv32i32(<vscale x 32 x ptr>, i32, <vscale x 32 x i1>, <vscale x 32 x i32>)
declare <vscale x 4 x half> @llvm.masked.gather.nxv4f16(<vscale x 4 x ptr>, i32, <vscale x 4 x i1>, <vscale x 4 x half>)
declare <vscale x 8 x half> @llvm.masked.gather.nxv8f16(<vscale x 8 x ptr>, i32, <vscale x 8 x i1>, <vscale x 8 x half>)
declare <vscale x 8 x bfloat> @llvm.masked.gather.nxv8bf16(<vscale x 8 x ptr>, i32, <vscale x 8 x i1>, <vscale x 8 x bfloat>)
declare <vscale x 2 x float> @llvm.masked.gather.nxv2f32(<vscale x 2 x ptr>, i32, <vscale x 2 x i1>, <vscale x 2 x float>)
declare <vscale x 8 x float> @llvm.masked.gather.nxv8f32(<vscale x 8 x ptr>, i32, <vscale x 8 x i1>, <vscale x 8 x float>)
declare <vscale x 4 x double> @llvm.masked.gather.nxv4f64(<vscale x 4 x ptr>, i32, <vscale x 4 x i1>, <vscale x 4 x double>)
|
