1 ; RUN: opt -mtriple=amdgcn-- -S -separate-const-offset-from-gep -reassociate-geps-verify-no-dead-code -gvn < %s | FileCheck -check-prefix=IR %s
3 target datalayout = "e-p:32:32-p1:64:64-p2:64:64-p3:32:32-p4:64:64-p5:32:32-p24:64:64-i64:64-v16:16-v24:32-v32:32-v48:64-v96:128-v192:256-v256:256-v512:512-v1024:1024-v2048:2048-n32:64"
5 @array = internal addrspace(2) constant [4096 x [32 x float]] zeroinitializer, align 4
7 ; IR-LABEL: @sum_of_array(
8 ; IR: [[BASE_PTR:%[a-zA-Z0-9]+]] = getelementptr inbounds [4096 x [32 x float]], [4096 x [32 x float]] addrspace(2)* @array, i64 0, i64 %{{[a-zA-Z0-9]+}}, i64 %{{[a-zA-Z0-9]+}}
9 ; IR: getelementptr float, float addrspace(2)* [[BASE_PTR]], i64 1
10 ; IR: getelementptr float, float addrspace(2)* [[BASE_PTR]], i64 32
11 ; IR: getelementptr float, float addrspace(2)* [[BASE_PTR]], i64 33
12 define void @sum_of_array(i32 %x, i32 %y, float addrspace(1)* nocapture %output) {
13 %tmp = sext i32 %y to i64
14 %tmp1 = sext i32 %x to i64
15 %tmp2 = getelementptr inbounds [4096 x [32 x float]], [4096 x [32 x float]] addrspace(2)* @array, i64 0, i64 %tmp1, i64 %tmp
16 %tmp4 = load float, float addrspace(2)* %tmp2, align 4
17 %tmp5 = fadd float %tmp4, 0.000000e+00
19 %tmp7 = sext i32 %tmp6 to i64
20 %tmp8 = getelementptr inbounds [4096 x [32 x float]], [4096 x [32 x float]] addrspace(2)* @array, i64 0, i64 %tmp1, i64 %tmp7
21 %tmp10 = load float, float addrspace(2)* %tmp8, align 4
22 %tmp11 = fadd float %tmp5, %tmp10
23 %tmp12 = add i32 %x, 1
24 %tmp13 = sext i32 %tmp12 to i64
25 %tmp14 = getelementptr inbounds [4096 x [32 x float]], [4096 x [32 x float]] addrspace(2)* @array, i64 0, i64 %tmp13, i64 %tmp
26 %tmp16 = load float, float addrspace(2)* %tmp14, align 4
27 %tmp17 = fadd float %tmp11, %tmp16
28 %tmp18 = getelementptr inbounds [4096 x [32 x float]], [4096 x [32 x float]] addrspace(2)* @array, i64 0, i64 %tmp13, i64 %tmp7
29 %tmp20 = load float, float addrspace(2)* %tmp18, align 4
30 %tmp21 = fadd float %tmp17, %tmp20
31 store float %tmp21, float addrspace(1)* %output, align 4
35 @array2 = internal addrspace(2) constant [4096 x [4 x float]] zeroinitializer, align 4
37 ; Some of the indices go over the maximum mubuf offset, so don't split them.
39 ; IR-LABEL: @sum_of_array_over_max_mubuf_offset(
40 ; IR: [[BASE_PTR:%[a-zA-Z0-9]+]] = getelementptr inbounds [4096 x [4 x float]], [4096 x [4 x float]] addrspace(2)* @array2, i64 0, i64 %{{[a-zA-Z0-9]+}}, i64 %{{[a-zA-Z0-9]+}}
41 ; IR: getelementptr float, float addrspace(2)* [[BASE_PTR]], i64 255
43 ; IR: getelementptr inbounds [4096 x [4 x float]], [4096 x [4 x float]] addrspace(2)* @array2, i64 0, i64 %{{[a-zA-Z0-9]+}}, i64 %{{[a-zA-Z0-9]+}}
44 ; IR: getelementptr inbounds [4096 x [4 x float]], [4096 x [4 x float]] addrspace(2)* @array2, i64 0, i64 %{{[a-zA-Z0-9]+}}, i64 %{{[a-zA-Z0-9]+}}
45 define void @sum_of_array_over_max_mubuf_offset(i32 %x, i32 %y, float addrspace(1)* nocapture %output) {
46 %tmp = sext i32 %y to i64
47 %tmp1 = sext i32 %x to i64
48 %tmp2 = getelementptr inbounds [4096 x [4 x float]], [4096 x [4 x float]] addrspace(2)* @array2, i64 0, i64 %tmp1, i64 %tmp
49 %tmp4 = load float, float addrspace(2)* %tmp2, align 4
50 %tmp5 = fadd float %tmp4, 0.000000e+00
51 %tmp6 = add i32 %y, 255
52 %tmp7 = sext i32 %tmp6 to i64
53 %tmp8 = getelementptr inbounds [4096 x [4 x float]], [4096 x [4 x float]] addrspace(2)* @array2, i64 0, i64 %tmp1, i64 %tmp7
54 %tmp10 = load float, float addrspace(2)* %tmp8, align 4
55 %tmp11 = fadd float %tmp5, %tmp10
56 %tmp12 = add i32 %x, 256
57 %tmp13 = sext i32 %tmp12 to i64
58 %tmp14 = getelementptr inbounds [4096 x [4 x float]], [4096 x [4 x float]] addrspace(2)* @array2, i64 0, i64 %tmp13, i64 %tmp
59 %tmp16 = load float, float addrspace(2)* %tmp14, align 4
60 %tmp17 = fadd float %tmp11, %tmp16
61 %tmp18 = getelementptr inbounds [4096 x [4 x float]], [4096 x [4 x float]] addrspace(2)* @array2, i64 0, i64 %tmp13, i64 %tmp7
62 %tmp20 = load float, float addrspace(2)* %tmp18, align 4
63 %tmp21 = fadd float %tmp17, %tmp20
64 store float %tmp21, float addrspace(1)* %output, align 4
69 @lds_array = internal addrspace(3) global [4096 x [4 x float]] undef, align 4
71 ; DS instructions have a larger immediate offset, so make sure these are OK.
72 ; IR-LABEL: @sum_of_lds_array_over_max_mubuf_offset(
73 ; IR: [[BASE_PTR:%[a-zA-Z0-9]+]] = getelementptr inbounds [4096 x [4 x float]], [4096 x [4 x float]] addrspace(3)* @lds_array, i32 0, i32 %{{[a-zA-Z0-9]+}}, i32 %{{[a-zA-Z0-9]+}}
74 ; IR: getelementptr float, float addrspace(3)* [[BASE_PTR]], i32 255
75 ; IR: getelementptr float, float addrspace(3)* [[BASE_PTR]], i32 16128
76 ; IR: getelementptr float, float addrspace(3)* [[BASE_PTR]], i32 16383
77 define void @sum_of_lds_array_over_max_mubuf_offset(i32 %x, i32 %y, float addrspace(1)* nocapture %output) {
78 %tmp2 = getelementptr inbounds [4096 x [4 x float]], [4096 x [4 x float]] addrspace(3)* @lds_array, i32 0, i32 %x, i32 %y
79 %tmp4 = load float, float addrspace(3)* %tmp2, align 4
80 %tmp5 = fadd float %tmp4, 0.000000e+00
81 %tmp6 = add i32 %y, 255
82 %tmp8 = getelementptr inbounds [4096 x [4 x float]], [4096 x [4 x float]] addrspace(3)* @lds_array, i32 0, i32 %x, i32 %tmp6
83 %tmp10 = load float, float addrspace(3)* %tmp8, align 4
84 %tmp11 = fadd float %tmp5, %tmp10
85 %tmp12 = add i32 %x, 4032
86 %tmp14 = getelementptr inbounds [4096 x [4 x float]], [4096 x [4 x float]] addrspace(3)* @lds_array, i32 0, i32 %tmp12, i32 %y
87 %tmp16 = load float, float addrspace(3)* %tmp14, align 4
88 %tmp17 = fadd float %tmp11, %tmp16
89 %tmp18 = getelementptr inbounds [4096 x [4 x float]], [4096 x [4 x float]] addrspace(3)* @lds_array, i32 0, i32 %tmp12, i32 %tmp6
90 %tmp20 = load float, float addrspace(3)* %tmp18, align 4
91 %tmp21 = fadd float %tmp17, %tmp20
92 store float %tmp21, float addrspace(1)* %output, align 4