1 ; RUN: llc -mtriple=amdgcn-amd-amdhsa -mcpu=kaveri -disable-promote-alloca-to-vector -amdgpu-enable-lower-module-lds=0 < %s | FileCheck -check-prefix=GCN %s
3 ; This shows that the amount LDS size estimate should try to not be
4 ; sensitive to the order of the LDS globals. This should try to
5 ; estimate the worst case padding behavior to avoid overallocating
8 ; These functions use the same amount of LDS, but the total, final
9 ; size changes depending on the visit order of first use.
11 ; The one with the suboptimal order resulting in extra padding exceeds
14 ; The padding estimate heuristic used by the promote alloca pass
15 ; is mostly determined by the order of the globals,
17 ; Raw usage = 1060 bytes
19 ; 292 + (4 pad) + 256 + (8 pad) + 512 = 1072
20 ; 512 + (0 pad) + 256 + (0 pad) + 292 = 1060
22 ; At default occupancy guess of 7, 2340 bytes available total.
24 ; 1280 need to be left to promote alloca
25 ; optimally packed, this requires
28 @lds0 = internal unnamed_addr addrspace(3) global [32 x <4 x i32>] undef, align 16
29 @lds2 = internal unnamed_addr addrspace(3) global [32 x i64] undef, align 8
30 @lds1 = internal unnamed_addr addrspace(3) global [73 x i32] undef, align 4
33 ; GCN-LABEL: {{^}}promote_alloca_size_order_0:
34 ; GCN: .amdhsa_group_segment_fixed_size 1060
35 define amdgpu_kernel void @promote_alloca_size_order_0(ptr addrspace(1) nocapture %out, ptr addrspace(1) nocapture %in, i32 %idx) #0 {
37 %stack = alloca [5 x i32], align 4, addrspace(5)
38 %tmp0 = load i32, ptr addrspace(1) %in, align 4
39 %arrayidx1 = getelementptr inbounds [5 x i32], ptr addrspace(5) %stack, i32 0, i32 %tmp0
40 store i32 4, ptr addrspace(5) %arrayidx1, align 4
41 %arrayidx2 = getelementptr inbounds i32, ptr addrspace(1) %in, i32 1
42 %tmp1 = load i32, ptr addrspace(1) %arrayidx2, align 4
43 %arrayidx3 = getelementptr inbounds [5 x i32], ptr addrspace(5) %stack, i32 0, i32 %tmp1
44 store i32 5, ptr addrspace(5) %arrayidx3, align 4
45 %tmp2 = load i32, ptr addrspace(5) %stack, align 4
46 store i32 %tmp2, ptr addrspace(1) %out, align 4
47 %arrayidx12 = getelementptr inbounds [5 x i32], ptr addrspace(5) %stack, i32 0, i32 1
48 %tmp3 = load i32, ptr addrspace(5) %arrayidx12
49 %arrayidx13 = getelementptr inbounds i32, ptr addrspace(1) %out, i32 1
50 store i32 %tmp3, ptr addrspace(1) %arrayidx13
52 %gep.lds1 = getelementptr inbounds [73 x i32], ptr addrspace(3) @lds1, i32 0, i32 %idx
53 store volatile i32 0, ptr addrspace(3) %gep.lds1, align 4
55 %gep.lds2 = getelementptr inbounds [32 x i64], ptr addrspace(3) @lds2, i32 0, i32 %idx
56 store volatile i64 0, ptr addrspace(3) %gep.lds2, align 8
58 %gep.lds0 = getelementptr inbounds [32 x <4 x i32>], ptr addrspace(3) @lds0, i32 0, i32 %idx
59 store volatile <4 x i32> zeroinitializer, ptr addrspace(3) %gep.lds0, align 16
64 ; GCN-LABEL: {{^}}promote_alloca_size_order_1:
65 ; GCN: .amdhsa_group_segment_fixed_size 1072
66 define amdgpu_kernel void @promote_alloca_size_order_1(ptr addrspace(1) nocapture %out, ptr addrspace(1) nocapture %in, i32 %idx) #0 {
68 %stack = alloca [5 x i32], align 4, addrspace(5)
69 %tmp0 = load i32, ptr addrspace(1) %in, align 4
70 %arrayidx1 = getelementptr inbounds [5 x i32], ptr addrspace(5) %stack, i32 0, i32 %tmp0
71 store i32 4, ptr addrspace(5) %arrayidx1, align 4
72 %arrayidx2 = getelementptr inbounds i32, ptr addrspace(1) %in, i32 1
73 %tmp1 = load i32, ptr addrspace(1) %arrayidx2, align 4
74 %arrayidx3 = getelementptr inbounds [5 x i32], ptr addrspace(5) %stack, i32 0, i32 %tmp1
75 store i32 5, ptr addrspace(5) %arrayidx3, align 4
76 %tmp2 = load i32, ptr addrspace(5) %stack, align 4
77 store i32 %tmp2, ptr addrspace(1) %out, align 4
78 %arrayidx12 = getelementptr inbounds [5 x i32], ptr addrspace(5) %stack, i32 0, i32 1
79 %tmp3 = load i32, ptr addrspace(5) %arrayidx12
80 %arrayidx13 = getelementptr inbounds i32, ptr addrspace(1) %out, i32 1
81 store i32 %tmp3, ptr addrspace(1) %arrayidx13
83 %gep.lds0 = getelementptr inbounds [32 x <4 x i32>], ptr addrspace(3) @lds0, i32 0, i32 %idx
84 store volatile <4 x i32> zeroinitializer, ptr addrspace(3) %gep.lds0, align 16
86 %gep.lds2 = getelementptr inbounds [32 x i64], ptr addrspace(3) @lds2, i32 0, i32 %idx
87 store volatile i64 0, ptr addrspace(3) %gep.lds2, align 8
89 %gep.lds1 = getelementptr inbounds [73 x i32], ptr addrspace(3) @lds1, i32 0, i32 %idx
90 store volatile i32 0, ptr 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: .amdhsa_group_segment_fixed_size 1060
103 define amdgpu_kernel void @promote_alloca_align_pad_guess_over_limit(ptr addrspace(1) nocapture %out, ptr addrspace(1) nocapture %in, i32 %idx) #0 {
105 %stack = alloca [5 x i32], align 4, addrspace(5)
106 %tmp0 = load i32, ptr addrspace(1) %in, align 4
107 %arrayidx1 = getelementptr inbounds [5 x i32], ptr addrspace(5) %stack, i32 0, i32 %tmp0
108 store i32 4, ptr addrspace(5) %arrayidx1, align 4
109 %arrayidx2 = getelementptr inbounds i32, ptr addrspace(1) %in, i32 1
110 %tmp1 = load i32, ptr addrspace(1) %arrayidx2, align 4
111 %arrayidx3 = getelementptr inbounds [5 x i32], ptr addrspace(5) %stack, i32 0, i32 %tmp1
112 store i32 5, ptr addrspace(5) %arrayidx3, align 4
113 %tmp2 = load i32, ptr addrspace(5) %stack, align 4
114 store i32 %tmp2, ptr addrspace(1) %out, align 4
115 %arrayidx12 = getelementptr inbounds [5 x i32], ptr addrspace(5) %stack, i32 0, i32 1
116 %tmp3 = load i32, ptr addrspace(5) %arrayidx12
117 %arrayidx13 = getelementptr inbounds i32, ptr addrspace(1) %out, i32 1
118 store i32 %tmp3, ptr addrspace(1) %arrayidx13
120 %gep.lds3 = getelementptr inbounds [13 x i32], ptr addrspace(3) @lds3, i32 0, i32 %idx
121 store volatile i32 0, ptr addrspace(3) %gep.lds3, align 4
123 %gep.lds4 = getelementptr inbounds [63 x <4 x i32>], ptr addrspace(3) @lds4, i32 0, i32 %idx
124 store volatile <4 x i32> zeroinitializer, ptr addrspace(3) %gep.lds4, align 16
129 attributes #0 = { nounwind "amdgpu-flat-work-group-size"="64,64" "amdgpu-waves-per-eu"="1,7" }
131 !llvm.module.flags = !{!0}
132 !0 = !{i32 1, !"amdhsa_code_object_version", i32 400}