• Home
  • Line#
  • Scopes#
  • Navigate#
  • Raw
  • Download
1; RUN: opt < %s -separate-const-offset-from-gep -slsr -gvn -S | FileCheck %s
2; RUN: llc < %s -march=nvptx64 -mcpu=sm_35 | FileCheck %s --check-prefix=PTX
3
4target datalayout = "e-i64:64-v16:16-v32:32-n16:32:64"
5target triple = "nvptx64-unknown-unknown"
6
7; arr[i + 5]
8; arr[i * 2 + 5]
9; arr[i * 3 + 5]
10; arr[i * 4 + 5]
11;
12;   => reassociate-geps
13;
14; *(&arr[i] + 5)
15; *(&arr[i * 2] + 5)
16; *(&arr[i * 3] + 5)
17; *(&arr[i * 4] + 5)
18;
19;   => slsr
20;
21; p1 = &arr[i]
22; *(p1 + 5)
23; p2 = p1 + i
24; *(p2 + 5)
25; p3 = p2 + i
26; *(p3 + 5)
27; p4 = p3 + i
28; *(p4 + 5)
29define void @slsr_after_reassociate_geps(float* %arr, i32 %i) {
30; CHECK-LABEL: @slsr_after_reassociate_geps(
31; PTX-LABEL: .visible .func slsr_after_reassociate_geps(
32; PTX: ld.param.u64 [[arr:%rd[0-9]+]], [slsr_after_reassociate_geps_param_0];
33; PTX: ld.param.u32 [[i:%r[0-9]+]], [slsr_after_reassociate_geps_param_1];
34  %i2 = shl nsw i32 %i, 1
35  %i3 = mul nsw i32 %i, 3
36  %i4 = shl nsw i32 %i, 2
37
38  %j1 = add nsw i32 %i, 5
39  %p1 = getelementptr inbounds float, float* %arr, i32 %j1
40; CHECK: [[b1:%[0-9]+]] = getelementptr float, float* %arr, i64 [[bump:%[0-9]+]]
41; PTX: mul.wide.s32 [[i4:%rd[0-9]+]], [[i]], 4;
42; PTX: add.s64 [[base1:%rd[0-9]+]], [[arr]], [[i4]];
43  %v1 = load float, float* %p1, align 4
44; PTX: ld.f32 {{%f[0-9]+}}, {{\[}}[[base1]]+20];
45  call void @foo(float %v1)
46
47  %j2 = add nsw i32 %i2, 5
48  %p2 = getelementptr inbounds float, float* %arr, i32 %j2
49; CHECK: [[b2:%[0-9]+]] = getelementptr float, float* [[b1]], i64 [[bump]]
50; PTX: add.s64 [[base2:%rd[0-9]+]], [[base1]], [[i4]];
51  %v2 = load float, float* %p2, align 4
52; PTX: ld.f32 {{%f[0-9]+}}, {{\[}}[[base2]]+20];
53  call void @foo(float %v2)
54
55  %j3 = add nsw i32 %i3, 5
56  %p3 = getelementptr inbounds float, float* %arr, i32 %j3
57; CHECK: [[b3:%[0-9]+]] = getelementptr float, float* [[b2]], i64 [[bump]]
58; PTX: add.s64 [[base3:%rd[0-9]+]], [[base2]], [[i4]];
59  %v3 = load float, float* %p3, align 4
60; PTX: ld.f32 {{%f[0-9]+}}, {{\[}}[[base3]]+20];
61  call void @foo(float %v3)
62
63  %j4 = add nsw i32 %i4, 5
64  %p4 = getelementptr inbounds float, float* %arr, i32 %j4
65; CHECK: [[b4:%[0-9]+]] = getelementptr float, float* [[b3]], i64 [[bump]]
66; PTX: add.s64 [[base4:%rd[0-9]+]], [[base3]], [[i4]];
67  %v4 = load float, float* %p4, align 4
68; PTX: ld.f32 {{%f[0-9]+}}, {{\[}}[[base4]]+20];
69  call void @foo(float %v4)
70
71  ret void
72}
73
74declare void @foo(float)
75