• Home
  • Line#
  • Scopes#
  • Navigate#
  • Raw
  • Download
1; RUN: llc < %s -march=nvptx -mcpu=sm_20 | FileCheck %s --check-prefix PTX
2; RUN: llc < %s -march=nvptx64 -mcpu=sm_20 | FileCheck %s --check-prefix PTX
3; RUN: opt < %s -S -nvptx-favor-non-generic -dce | FileCheck %s --check-prefix IR
4
5@array = internal addrspace(3) global [10 x float] zeroinitializer, align 4
6@scalar = internal addrspace(3) global float 0.000000e+00, align 4
7
8; Verifies nvptx-favor-non-generic correctly optimizes generic address space
9; usage to non-generic address space usage for the patterns we claim to handle:
10; 1. load cast
11; 2. store cast
12; 3. load gep cast
13; 4. store gep cast
14; gep and cast can be an instruction or a constant expression. This function
15; tries all possible combinations.
16define float @ld_st_shared_f32(i32 %i, float %v) {
17; IR-LABEL: @ld_st_shared_f32
18; IR-NOT: addrspacecast
19; PTX-LABEL: ld_st_shared_f32(
20  ; load cast
21  %1 = load float* addrspacecast (float addrspace(3)* @scalar to float*), align 4
22; PTX: ld.shared.f32 %f{{[0-9]+}}, [scalar];
23  ; store cast
24  store float %v, float* addrspacecast (float addrspace(3)* @scalar to float*), align 4
25; PTX: st.shared.f32 [scalar], %f{{[0-9]+}};
26  ; use syncthreads to disable optimizations across components
27  call void @llvm.cuda.syncthreads()
28; PTX: bar.sync 0;
29
30  ; cast; load
31  %2 = addrspacecast float addrspace(3)* @scalar to float*
32  %3 = load float* %2, align 4
33; PTX: ld.shared.f32 %f{{[0-9]+}}, [scalar];
34  ; cast; store
35  store float %v, float* %2, align 4
36; PTX: st.shared.f32 [scalar], %f{{[0-9]+}};
37  call void @llvm.cuda.syncthreads()
38; PTX: bar.sync 0;
39
40  ; load gep cast
41  %4 = load float* getelementptr inbounds ([10 x float]* addrspacecast ([10 x float] addrspace(3)* @array to [10 x float]*), i32 0, i32 5), align 4
42; PTX: ld.shared.f32 %f{{[0-9]+}}, [array+20];
43  ; store gep cast
44  store float %v, float* getelementptr inbounds ([10 x float]* addrspacecast ([10 x float] addrspace(3)* @array to [10 x float]*), i32 0, i32 5), align 4
45; PTX: st.shared.f32 [array+20], %f{{[0-9]+}};
46  call void @llvm.cuda.syncthreads()
47; PTX: bar.sync 0;
48
49  ; gep cast; load
50  %5 = getelementptr inbounds [10 x float]* addrspacecast ([10 x float] addrspace(3)* @array to [10 x float]*), i32 0, i32 5
51  %6 = load float* %5, align 4
52; PTX: ld.shared.f32 %f{{[0-9]+}}, [array+20];
53  ; gep cast; store
54  store float %v, float* %5, align 4
55; PTX: st.shared.f32 [array+20], %f{{[0-9]+}};
56  call void @llvm.cuda.syncthreads()
57; PTX: bar.sync 0;
58
59  ; cast; gep; load
60  %7 = addrspacecast [10 x float] addrspace(3)* @array to [10 x float]*
61  %8 = getelementptr inbounds [10 x float]* %7, i32 0, i32 %i
62  %9 = load float* %8, align 4
63; PTX: ld.shared.f32 %f{{[0-9]+}}, [%{{(r|rl|rd)[0-9]+}}];
64  ; cast; gep; store
65  store float %v, float* %8, align 4
66; PTX: st.shared.f32 [%{{(r|rl|rd)[0-9]+}}], %f{{[0-9]+}};
67  call void @llvm.cuda.syncthreads()
68; PTX: bar.sync 0;
69
70  %sum2 = fadd float %1, %3
71  %sum3 = fadd float %sum2, %4
72  %sum4 = fadd float %sum3, %6
73  %sum5 = fadd float %sum4, %9
74  ret float %sum5
75}
76
77; When hoisting an addrspacecast between different pointer types, replace the
78; addrspacecast with a bitcast.
79define i32 @ld_int_from_float() {
80; IR-LABEL: @ld_int_from_float
81; IR: load i32 addrspace(3)* bitcast (float addrspace(3)* @scalar to i32 addrspace(3)*)
82; PTX-LABEL: ld_int_from_float(
83; PTX: ld.shared.u{{(32|64)}}
84  %1 = load i32* addrspacecast(float addrspace(3)* @scalar to i32*), align 4
85  ret i32 %1
86}
87
88declare void @llvm.cuda.syncthreads() #3
89
90attributes #3 = { noduplicate nounwind }
91
92