1 ; RUN: llc -mtriple=amdgcn-amd-amdhsa -mcpu=kaveri -mattr=-code-object-v3 -disable-promote-alloca-to-vector < %s | FileCheck -check-prefix=GCN %s
3 ; This shows that the amount of LDS estimate is sensitive to the order
6 ; Both of these functions use the same amount of LDS, but the total
7 ; changes depending on the visit order of first use.
9 ; The one with the suboptimal order resulting in extra padding exceeds
12 ; The padding estimate heuristic used by the promote alloca pass
13 ; is mostly determined by the order of the globals,
15 ; Raw usage = 1060 bytes
17 ; 292 + (4 pad) + 256 + (8 pad) + 512 = 1072
18 ; 512 + (0 pad) + 256 + (0 pad) + 292 = 1060
20 ; At default occupancy guess of 7, 2340 bytes available total.
22 ; 1280 need to be left to promote alloca
23 ; optimally packed, this requires
26 @lds0 = internal unnamed_addr addrspace(3) global [32 x <4 x i32>] undef, align 16
27 @lds2 = internal unnamed_addr addrspace(3) global [32 x i64] undef, align 8
28 @lds1 = internal unnamed_addr addrspace(3) global [73 x i32] undef, align 4
31 ; GCN-LABEL: {{^}}promote_alloca_size_order_0:
32 ; GCN: workgroup_group_segment_byte_size = 2340
33 define amdgpu_kernel void @promote_alloca_size_order_0(i32 addrspace(1)* nocapture %out, i32 addrspace(1)* nocapture %in, i32 %idx) #0 {
35 %stack = alloca [5 x i32], align 4, addrspace(5)
36 %tmp0 = load i32, i32 addrspace(1)* %in, align 4
37 %arrayidx1 = getelementptr inbounds [5 x i32], [5 x i32] addrspace(5)* %stack, i32 0, i32 %tmp0
38 store i32 4, i32 addrspace(5)* %arrayidx1, align 4
39 %arrayidx2 = getelementptr inbounds i32, i32 addrspace(1)* %in, i32 1
40 %tmp1 = load i32, i32 addrspace(1)* %arrayidx2, align 4
41 %arrayidx3 = getelementptr inbounds [5 x i32], [5 x i32] addrspace(5)* %stack, i32 0, i32 %tmp1
42 store i32 5, i32 addrspace(5)* %arrayidx3, align 4
43 %arrayidx10 = getelementptr inbounds [5 x i32], [5 x i32] addrspace(5)* %stack, i32 0, i32 0
44 %tmp2 = load i32, i32 addrspace(5)* %arrayidx10, align 4
45 store i32 %tmp2, i32 addrspace(1)* %out, align 4
46 %arrayidx12 = getelementptr inbounds [5 x i32], [5 x i32] addrspace(5)* %stack, i32 0, i32 1
47 %tmp3 = load i32, i32 addrspace(5)* %arrayidx12
48 %arrayidx13 = getelementptr inbounds i32, i32 addrspace(1)* %out, i32 1
49 store i32 %tmp3, i32 addrspace(1)* %arrayidx13
51 %gep.lds1 = getelementptr inbounds [73 x i32], [73 x i32] addrspace(3)* @lds1, i32 0, i32 %idx
52 store volatile i32 0, i32 addrspace(3)* %gep.lds1, align 4
54 %gep.lds2 = getelementptr inbounds [32 x i64], [32 x i64] addrspace(3)* @lds2, i32 0, i32 %idx
55 store volatile i64 0, i64 addrspace(3)* %gep.lds2, align 8
57 %gep.lds0 = getelementptr inbounds [32 x <4 x i32>], [32 x <4 x i32>] addrspace(3)* @lds0, i32 0, i32 %idx
58 store volatile <4 x i32> zeroinitializer, <4 x i32> addrspace(3)* %gep.lds0, align 16
63 ; GCN-LABEL: {{^}}promote_alloca_size_order_1:
64 ; GCN: workgroup_group_segment_byte_size = 2352
65 define amdgpu_kernel void @promote_alloca_size_order_1(i32 addrspace(1)* nocapture %out, i32 addrspace(1)* nocapture %in, i32 %idx) #0 {
67 %stack = alloca [5 x i32], align 4, addrspace(5)
68 %tmp0 = load i32, i32 addrspace(1)* %in, align 4
69 %arrayidx1 = getelementptr inbounds [5 x i32], [5 x i32] addrspace(5)* %stack, i32 0, i32 %tmp0
70 store i32 4, i32 addrspace(5)* %arrayidx1, align 4
71 %arrayidx2 = getelementptr inbounds i32, i32 addrspace(1)* %in, i32 1
72 %tmp1 = load i32, i32 addrspace(1)* %arrayidx2, align 4
73 %arrayidx3 = getelementptr inbounds [5 x i32], [5 x i32] addrspace(5)* %stack, i32 0, i32 %tmp1
74 store i32 5, i32 addrspace(5)* %arrayidx3, align 4
75 %arrayidx10 = getelementptr inbounds [5 x i32], [5 x i32] addrspace(5)* %stack, i32 0, i32 0
76 %tmp2 = load i32, i32 addrspace(5)* %arrayidx10, align 4
77 store i32 %tmp2, i32 addrspace(1)* %out, align 4
78 %arrayidx12 = getelementptr inbounds [5 x i32], [5 x i32] addrspace(5)* %stack, i32 0, i32 1
79 %tmp3 = load i32, i32 addrspace(5)* %arrayidx12
80 %arrayidx13 = getelementptr inbounds i32, i32 addrspace(1)* %out, i32 1
81 store i32 %tmp3, i32 addrspace(1)* %arrayidx13
83 %gep.lds0 = getelementptr inbounds [32 x <4 x i32>], [32 x <4 x i32>] addrspace(3)* @lds0, i32 0, i32 %idx
84 store volatile <4 x i32> zeroinitializer, <4 x i32> addrspace(3)* %gep.lds0, align 16
86 %gep.lds2 = getelementptr inbounds [32 x i64], [32 x i64] addrspace(3)* @lds2, i32 0, i32 %idx
87 store volatile i64 0, i64 addrspace(3)* %gep.lds2, align 8
89 %gep.lds1 = getelementptr inbounds [73 x i32], [73 x i32] addrspace(3)* @lds1, i32 0, i32 %idx
90 store volatile i32 0, i32 addrspace(3)* %gep.lds1, align 4
95 @lds3 = internal unnamed_addr addrspace(3) global [13 x i32] undef, align 4
96 @lds4 = internal unnamed_addr addrspace(3) global [63 x <4 x i32>] undef, align 16
98 ; The guess from the alignment padding pushes this over the determined
99 ; size limit, so it isn't promoted
101 ; GCN-LABEL: {{^}}promote_alloca_align_pad_guess_over_limit:
102 ; GCN: workgroup_group_segment_byte_size = 1060
103 define amdgpu_kernel void @promote_alloca_align_pad_guess_over_limit(i32 addrspace(1)* nocapture %out, i32 addrspace(1)* nocapture %in, i32 %idx) #0 {
105 %stack = alloca [5 x i32], align 4, addrspace(5)
106 %tmp0 = load i32, i32 addrspace(1)* %in, align 4
107 %arrayidx1 = getelementptr inbounds [5 x i32], [5 x i32] addrspace(5)* %stack, i32 0, i32 %tmp0
108 store i32 4, i32 addrspace(5)* %arrayidx1, align 4
109 %arrayidx2 = getelementptr inbounds i32, i32 addrspace(1)* %in, i32 1
110 %tmp1 = load i32, i32 addrspace(1)* %arrayidx2, align 4
111 %arrayidx3 = getelementptr inbounds [5 x i32], [5 x i32] addrspace(5)* %stack, i32 0, i32 %tmp1
112 store i32 5, i32 addrspace(5)* %arrayidx3, align 4
113 %arrayidx10 = getelementptr inbounds [5 x i32], [5 x i32] addrspace(5)* %stack, i32 0, i32 0
114 %tmp2 = load i32, i32 addrspace(5)* %arrayidx10, align 4
115 store i32 %tmp2, i32 addrspace(1)* %out, align 4
116 %arrayidx12 = getelementptr inbounds [5 x i32], [5 x i32] addrspace(5)* %stack, i32 0, i32 1
117 %tmp3 = load i32, i32 addrspace(5)* %arrayidx12
118 %arrayidx13 = getelementptr inbounds i32, i32 addrspace(1)* %out, i32 1
119 store i32 %tmp3, i32 addrspace(1)* %arrayidx13
121 %gep.lds3 = getelementptr inbounds [13 x i32], [13 x i32] addrspace(3)* @lds3, i32 0, i32 %idx
122 store volatile i32 0, i32 addrspace(3)* %gep.lds3, align 4
124 %gep.lds4 = getelementptr inbounds [63 x <4 x i32>], [63 x <4 x i32>] addrspace(3)* @lds4, i32 0, i32 %idx
125 store volatile <4 x i32> zeroinitializer, <4 x i32> addrspace(3)* %gep.lds4, align 16
130 attributes #0 = { nounwind "amdgpu-flat-work-group-size"="64,64" "amdgpu-waves-per-eu"="1,7" }