xref: /llvm-project/llvm/test/CodeGen/AMDGPU/stack-size-overflow.ll (revision bc6955f18ced3ca89d49bc28eeb58cd6d367e136)
1; RUN: not llc -mtriple=amdgcn-- < %s 2>&1 | FileCheck -check-prefix=ERROR %s
2; RUN: not llc -mtriple=amdgcn-- < %s | FileCheck -check-prefix=GCN %s
3
4declare void @llvm.memset.p5.i32(ptr addrspace(5) nocapture, i8, i32, i32, i1) #1
5
6; ERROR: error: <unknown>:0:0: stack frame size (131064) exceeds limit (131056) in function 'stack_size_limit_wave64'
7; GCN: ; ScratchSize: 131064
8define amdgpu_kernel void @stack_size_limit_wave64() #0 {
9entry:
10  %alloca = alloca [131057 x i8], align 1, addrspace(5)
11  call void @llvm.memset.p5.i32(ptr addrspace(5) %alloca, i8 9, i32 131057, i32 1, i1 true)
12  ret void
13}
14
15; ERROR: error: <unknown>:0:0: stack frame size (262120) exceeds limit (262112) in function 'stack_size_limit_wave32'
16; GCN: ; ScratchSize: 262120
17define amdgpu_kernel void @stack_size_limit_wave32() #1 {
18entry:
19  %alloca = alloca [262113 x i8], align 1, addrspace(5)
20  call void @llvm.memset.p5.i32(ptr addrspace(5) %alloca, i8 9, i32 262113, i32 1, i1 true)
21  ret void
22}
23
24; ERROR-NOT: error:
25; GCN: ; ScratchSize: 131056
26define amdgpu_kernel void @max_stack_size_wave64() #0 {
27entry:
28  %alloca = alloca [131052 x i8], align 1, addrspace(5)
29  call void @llvm.memset.p5.i32(ptr addrspace(5) %alloca, i8 9, i32 131052, i32 1, i1 true)
30  ret void
31}
32
33; ERROR-NOT: error:
34; GCN: ; ScratchSize: 262112
35define amdgpu_kernel void @max_stack_size_wave32() #1 {
36entry:
37  %alloca = alloca [262108 x i8], align 1, addrspace(5)
38  call void @llvm.memset.p5.i32(ptr addrspace(5) %alloca, i8 9, i32 262108, i32 1, i1 true)
39  ret void
40}
41
42attributes #0 = { "target-cpu" = "gfx900" }
43attributes #1 = { "target-cpu" = "gfx1010" }
44