1
; This provides optimized implementations of vload2/3/4/8/16 for 32-bit int/uint
2
; The address spaces get mapped to data types in target-specific usages
4
define <2 x i32> @__clc_vload2_i32__addr1(i32 addrspace(1)* nocapture %addr) nounwind readonly alwaysinline {
5
%1 = bitcast i32 addrspace(1)* %addr to <2 x i32> addrspace(1)*
6
%2 = load <2 x i32> addrspace(1)* %1, align 4, !tbaa !3
10
define <3 x i32> @__clc_vload3_i32__addr1(i32 addrspace(1)* nocapture %addr) nounwind readonly alwaysinline {
11
%1 = bitcast i32 addrspace(1)* %addr to <3 x i32> addrspace(1)*
12
%2 = load <3 x i32> addrspace(1)* %1, align 4, !tbaa !3
16
define <4 x i32> @__clc_vload4_i32__addr1(i32 addrspace(1)* nocapture %addr) nounwind readonly alwaysinline {
17
%1 = bitcast i32 addrspace(1)* %addr to <4 x i32> addrspace(1)*
18
%2 = load <4 x i32> addrspace(1)* %1, align 4, !tbaa !3
22
define <8 x i32> @__clc_vload8_i32__addr1(i32 addrspace(1)* nocapture %addr) nounwind readonly alwaysinline {
23
%1 = bitcast i32 addrspace(1)* %addr to <8 x i32> addrspace(1)*
24
%2 = load <8 x i32> addrspace(1)* %1, align 4, !tbaa !3
28
define <16 x i32> @__clc_vload16_i32__addr1(i32 addrspace(1)* nocapture %addr) nounwind readonly alwaysinline {
29
%1 = bitcast i32 addrspace(1)* %addr to <16 x i32> addrspace(1)*
30
%2 = load <16 x i32> addrspace(1)* %1, align 4, !tbaa !3
34
define <2 x i32> @__clc_vload2_i32__addr2(i32 addrspace(2)* nocapture %addr) nounwind readonly alwaysinline {
35
%1 = bitcast i32 addrspace(2)* %addr to <2 x i32> addrspace(2)*
36
%2 = load <2 x i32> addrspace(2)* %1, align 4, !tbaa !3
40
define <3 x i32> @__clc_vload3_i32__addr2(i32 addrspace(2)* nocapture %addr) nounwind readonly alwaysinline {
41
%1 = bitcast i32 addrspace(2)* %addr to <3 x i32> addrspace(2)*
42
%2 = load <3 x i32> addrspace(2)* %1, align 4, !tbaa !3
46
define <4 x i32> @__clc_vload4_i32__addr2(i32 addrspace(2)* nocapture %addr) nounwind readonly alwaysinline {
47
%1 = bitcast i32 addrspace(2)* %addr to <4 x i32> addrspace(2)*
48
%2 = load <4 x i32> addrspace(2)* %1, align 4, !tbaa !3
52
define <8 x i32> @__clc_vload8_i32__addr2(i32 addrspace(2)* nocapture %addr) nounwind readonly alwaysinline {
53
%1 = bitcast i32 addrspace(2)* %addr to <8 x i32> addrspace(2)*
54
%2 = load <8 x i32> addrspace(2)* %1, align 4, !tbaa !3
58
define <16 x i32> @__clc_vload16_i32__addr2(i32 addrspace(2)* nocapture %addr) nounwind readonly alwaysinline {
59
%1 = bitcast i32 addrspace(2)* %addr to <16 x i32> addrspace(2)*
60
%2 = load <16 x i32> addrspace(2)* %1, align 4, !tbaa !3
64
define <2 x i32> @__clc_vload2_i32__addr3(i32 addrspace(3)* nocapture %addr) nounwind readonly alwaysinline {
65
%1 = bitcast i32 addrspace(3)* %addr to <2 x i32> addrspace(3)*
66
%2 = load <2 x i32> addrspace(3)* %1, align 4, !tbaa !3
70
define <3 x i32> @__clc_vload3_i32__addr3(i32 addrspace(3)* nocapture %addr) nounwind readonly alwaysinline {
71
%1 = bitcast i32 addrspace(3)* %addr to <3 x i32> addrspace(3)*
72
%2 = load <3 x i32> addrspace(3)* %1, align 4, !tbaa !3
76
define <4 x i32> @__clc_vload4_i32__addr3(i32 addrspace(3)* nocapture %addr) nounwind readonly alwaysinline {
77
%1 = bitcast i32 addrspace(3)* %addr to <4 x i32> addrspace(3)*
78
%2 = load <4 x i32> addrspace(3)* %1, align 4, !tbaa !3
82
define <8 x i32> @__clc_vload8_i32__addr3(i32 addrspace(3)* nocapture %addr) nounwind readonly alwaysinline {
83
%1 = bitcast i32 addrspace(3)* %addr to <8 x i32> addrspace(3)*
84
%2 = load <8 x i32> addrspace(3)* %1, align 4, !tbaa !3
88
define <16 x i32> @__clc_vload16_i32__addr3(i32 addrspace(3)* nocapture %addr) nounwind readonly alwaysinline {
89
%1 = bitcast i32 addrspace(3)* %addr to <16 x i32> addrspace(3)*
90
%2 = load <16 x i32> addrspace(3)* %1, align 4, !tbaa !3
94
define <2 x i32> @__clc_vload2_i32__addr4(i32 addrspace(4)* nocapture %addr) nounwind readonly alwaysinline {
95
%1 = bitcast i32 addrspace(4)* %addr to <2 x i32> addrspace(4)*
96
%2 = load <2 x i32> addrspace(4)* %1, align 4, !tbaa !3
100
define <3 x i32> @__clc_vload3_i32__addr4(i32 addrspace(4)* nocapture %addr) nounwind readonly alwaysinline {
101
%1 = bitcast i32 addrspace(4)* %addr to <3 x i32> addrspace(4)*
102
%2 = load <3 x i32> addrspace(4)* %1, align 4, !tbaa !3
106
define <4 x i32> @__clc_vload4_i32__addr4(i32 addrspace(4)* nocapture %addr) nounwind readonly alwaysinline {
107
%1 = bitcast i32 addrspace(4)* %addr to <4 x i32> addrspace(4)*
108
%2 = load <4 x i32> addrspace(4)* %1, align 4, !tbaa !3
112
define <8 x i32> @__clc_vload8_i32__addr4(i32 addrspace(4)* nocapture %addr) nounwind readonly alwaysinline {
113
%1 = bitcast i32 addrspace(4)* %addr to <8 x i32> addrspace(4)*
114
%2 = load <8 x i32> addrspace(4)* %1, align 4, !tbaa !3
118
define <16 x i32> @__clc_vload16_i32__addr4(i32 addrspace(4)* nocapture %addr) nounwind readonly alwaysinline {
119
%1 = bitcast i32 addrspace(4)* %addr to <16 x i32> addrspace(4)*
120
%2 = load <16 x i32> addrspace(4)* %1, align 4, !tbaa !3
124
!1 = metadata !{metadata !"char", metadata !5}
125
!2 = metadata !{metadata !"short", metadata !5}
126
!3 = metadata !{metadata !"int", metadata !5}
127
!4 = metadata !{metadata !"long", metadata !5}
128
!5 = metadata !{metadata !"omnipotent char", metadata !6}
129
!6 = metadata !{metadata !"Simple C/C++ TBAA"}