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
|
;=========================== begin_copyright_notice ============================
;
; Copyright (C) 2025 Intel Corporation
;
; SPDX-License-Identifier: MIT
;
;============================ end_copyright_notice =============================
; This is to test the optimization. Normally, those tests should use struct as the type for
; loaded value. However, as vector has been extensively optmized. Load combining specially
; generates vector instead of struct for those tests, thus result in less instructions.
; REQUIRES: llvm-14-plus, regkeys
; RUN: igc_opt --opaque-pointers -S -inputocl -igc-ldstcombine -regkey=EnableLdStCombine=5 -platformbmg %s \
; RUN: | FileCheck %s
target datalayout = "e-p:64:64:64-i1:8:8-i8:8:8-i16:16:16-i32:32:32-i64:64:64-f32:32:32-f64:64:64-v16:16:16-v24:32:32-v32:32:32-v48:64:64-v64:64:64-v96:128:128-v128:128:128-v192:256:256-v256:256:256-v512:512:512-v1024:1024:1024-n8:16:32"
target triple = "spir64-unknown-unknown"
; CHECK-LABEL: define spir_kernel void @test_vector_creation
; Function Attrs: convergent nounwind
define spir_kernel void @test_vector_creation(i8 addrspace(1)* %src, <4 x i8> addrspace(1)* %result) #0 {
entry:
; case 0: 4 load i8 --> load i32 --> bitcast <4 x i8> --> store
; handle insertelement
; CHECK-LABEL: c0.base
; CHECK: [[T0_0:%.*]] = call i32 @llvm.genx.GenISA.PredicatedLoad.i32.p1.i32(ptr addrspace(1) %{{.*}}, i32 84148994)
; CHECK: {{.*}} = bitcast i32 [[T0_0]] to <4 x i8>
%c0.base = getelementptr inbounds i8, i8 addrspace(1)* %src, i64 0
%c0.0 = call i8 @llvm.genx.GenISA.PredicatedLoad.i8.p1i8.i8(i8 addrspace(1)* %c0.base, i64 1, i1 true, i8 2)
%c0.arrayidx1 = getelementptr inbounds i8, i8 addrspace(1)* %c0.base, i64 1
%c0.1 = call i8 @llvm.genx.GenISA.PredicatedLoad.i8.p1i8.i8(i8 addrspace(1)* %c0.arrayidx1, i64 1, i1 true, i8 3)
%c0.arrayidx3 = getelementptr inbounds i8, i8 addrspace(1)* %c0.base, i64 2
%c0.2 = call i8 @llvm.genx.GenISA.PredicatedLoad.i8.p1i8.i8(i8 addrspace(1)* %c0.arrayidx3, i64 1, i1 true, i8 4)
%c0.arrayidx5 = getelementptr inbounds i8, i8 addrspace(1)* %c0.base, i64 3
%c0.3 = call i8 @llvm.genx.GenISA.PredicatedLoad.i8.p1i8.i8(i8 addrspace(1)* %c0.arrayidx5, i64 1, i1 true, i8 5)
%c0.v.0 = insertelement <4 x i8> undef, i8 %c0.0, i32 0
%c0.v.1 = insertelement <4 x i8> %c0.v.0, i8 %c0.1, i32 1
%c0.v.2 = insertelement <4 x i8> %c0.v.1, i8 %c0.2, i32 2
%c0.v.3 = insertelement <4 x i8> %c0.v.2, i8 %c0.3, i32 3
%c0.dst = getelementptr inbounds <4 x i8>, <4 x i8> addrspace(1)* %result, i64 0
store <4 x i8> %c0.v.3, <4 x i8> addrspace(1)* %c0.dst, align 4
; case 1: [i8, i8, <2 x i8>] --> load i32
; handle insertelement and shufflevector
; CHECK-LABEL: c1.base
; CHECK: [[T1_0:%.*]] = call i32 @llvm.genx.GenISA.PredicatedLoad.i32.p1.i32(ptr addrspace(1) %{{.*}}, i32 84148994)
; CHECK: {{.*}} = bitcast i32 [[T1_0]] to <4 x i8>
%c1.base = getelementptr inbounds i8, i8 addrspace(1)* %src, i64 16
%c1.0 = call i8 @llvm.genx.GenISA.PredicatedLoad.i8.p1i8.i8(i8 addrspace(1)* %c1.base, i64 1, i1 true, i8 2)
%c1.arrayidx1 = getelementptr inbounds i8, i8 addrspace(1)* %c1.base, i64 1
%c1.1 = call i8 @llvm.genx.GenISA.PredicatedLoad.i8.p1i8.i8(i8 addrspace(1)* %c1.arrayidx1, i64 1, i1 true, i8 3)
%c1.arrayidx3 = getelementptr inbounds i8, i8 addrspace(1)* %c1.base, i64 2
%c1.arrayidx3.0 = bitcast i8 addrspace(1)* %c1.arrayidx3 to <2 x i8> addrspace(1)*
%c1.2 = call <2 x i8> @llvm.genx.GenISA.PredicatedLoad.v2i8.p1v2i8.v2i8(<2 x i8> addrspace(1)* %c1.arrayidx3.0, i64 1, i1 true, <2 x i8> <i8 4, i8 5>)
%c1.2.0 = shufflevector <2 x i8> %c1.2, <2 x i8> undef, <4 x i32> <i32 0, i32 1, i32 undef, i32 undef>
%c1.v.0 = insertelement <4 x i8> undef, i8 %c1.0, i32 0
%c1.v.1 = insertelement <4 x i8> %c1.v.0, i8 %c1.1, i32 1
%c1.v.2 = shufflevector <4 x i8> %c1.v.1, <4 x i8> %c1.2.0, <4 x i32> <i32 0, i32 1, i32 4, i32 5>
%c1.dst = getelementptr inbounds <4 x i8>, <4 x i8> addrspace(1)* %result, i64 2
store <4 x i8> %c1.v.2, <4 x i8> addrspace(1)* %c1.dst, align 4
; case 2: [i8, <2 x i8>, i8] --> load i32
; handle extractelement and insertelement
; CHECK-LABEL: c2.base
; CHECK: [[T2_0:%.*]] = call i32 @llvm.genx.GenISA.PredicatedLoad.i32.p1.i32(ptr addrspace(1) %{{.*}}, i32 84148994)
; CHECK: {{.*}} = bitcast i32 [[T2_0]] to <4 x i8>
%c2.base = getelementptr inbounds i8, i8 addrspace(1)* %src, i64 32
%c2.0 = call i8 @llvm.genx.GenISA.PredicatedLoad.i8.p1i8.i8(i8 addrspace(1)* %c2.base, i64 1, i1 true, i8 2)
%c2.arrayidx1 = getelementptr inbounds i8, i8 addrspace(1)* %c2.base, i64 1
%c2.arrayidx1.0 = bitcast i8 addrspace(1)* %c2.arrayidx1 to <2 x i8> addrspace(1)*
%c2.1 = call <2 x i8> @llvm.genx.GenISA.PredicatedLoad.v2i8.p1v2i8.v2i8(<2 x i8> addrspace(1)* %c2.arrayidx1.0, i64 1, i1 true, <2 x i8> <i8 3, i8 4>)
%c2.1.0 = extractelement <2 x i8> %c2.1, i32 0
%c2.1.1 = extractelement <2 x i8> %c2.1, i32 1
%c2.arrayidx3 = getelementptr inbounds i8, i8 addrspace(1)* %c2.base, i64 3
%c2.2 = call i8 @llvm.genx.GenISA.PredicatedLoad.i8.p1i8.i8(i8 addrspace(1)* %c2.arrayidx3, i64 1, i1 true, i8 5)
%c2.v.0 = insertelement <4 x i8> undef, i8 %c2.0, i32 0
%c2.v.1 = insertelement <4 x i8> %c2.v.0, i8 %c2.1.0, i32 1
%c2.v.2 = insertelement <4 x i8> %c2.v.1, i8 %c2.1.1, i32 2
%c2.v.3 = insertelement <4 x i8> %c2.v.2, i8 %c2.2, i32 3
%c2.dst = getelementptr inbounds <4 x i8>, <4 x i8> addrspace(1)* %result, i64 4
store <4 x i8> %c2.v.3, <4 x i8> addrspace(1)* %c2.dst, align 4
; case 3 : [ <2 x i8> <2 x i8> ] --> i32
; handle shufflevector
; CHECK-LABEL: c3.base
; CHECK: [[T3_0:%.*]] = call i32 @llvm.genx.GenISA.PredicatedLoad.i32.p1.i32(ptr addrspace(1) %{{.*}}, i32 84148994)
; CHECK: {{.*}} = bitcast i32 [[T3_0]] to <4 x i8>
; CHECK: ret void
%c3.base = getelementptr inbounds i8, i8 addrspace(1)* %src, i64 48
%c3.base.0 = bitcast i8 addrspace(1)* %c3.base to <2 x i8> addrspace(1)*
%c3.arrayidx1 = getelementptr inbounds <2 x i8>, <2 x i8> addrspace(1)* %c3.base.0, i64 0
%c3.0 = call <2 x i8> @llvm.genx.GenISA.PredicatedLoad.v2i8.p1v2i8.v2i8(<2 x i8> addrspace(1)* %c3.arrayidx1, i64 1, i1 true, <2 x i8> <i8 2, i8 3>)
%c3.arrayidx3 = getelementptr inbounds <2 x i8>, <2 x i8> addrspace(1)* %c3.base.0, i64 1
%c3.1 = call <2 x i8> @llvm.genx.GenISA.PredicatedLoad.v2i8.p1v2i8.v2i8(<2 x i8> addrspace(1)* %c3.arrayidx3, i64 1, i1 true, <2 x i8> <i8 4, i8 5>)
%c3.v = shufflevector <2 x i8> %c3.0, <2 x i8> %c3.1, <4 x i32> <i32 0, i32 1, i32 2, i32 3>
%c3.dst = getelementptr inbounds <4 x i8>, <4 x i8> addrspace(1)* %result, i64 10
store <4 x i8> %c3.v, <4 x i8> addrspace(1)* %c3.dst, align 4
ret void
}
; Function Attrs: nounwind readonly
declare i8 @llvm.genx.GenISA.PredicatedLoad.i8.p1i8.i8(i8 addrspace(1)*, i64, i1, i8) #1
; Function Attrs: nounwind readonly
declare <2 x i8> @llvm.genx.GenISA.PredicatedLoad.v2i8.p1v2i8.v2i8(<2 x i8> addrspace(1)*, i64, i1, <2 x i8>) #1
attributes #0 = { convergent nounwind "less-precise-fpmad"="true" "null-pointer-is-valid"="true" }
attributes #1 = { nounwind readonly }
|
