llvm/test/CodeGen/AMDGPU/recursion.ll

   1 ; RUN: sed 's/CODE_OBJECT_VERSION/400/g' %s | llc -mtriple=amdgcn-amd-amdhsa -mcpu=gfx900 -verify-machineinstrs | FileCheck %s
   2 ; RUN: sed 's/CODE_OBJECT_VERSION/500/g' %s | llc -mtriple=amdgcn-amd-amdhsa -mcpu=gfx900 -verify-machineinstrs | FileCheck -check-prefixes=V5 %s
   3 ; RUN: sed 's/CODE_OBJECT_VERSION/600/g' %s | llc -mtriple=amdgcn-amd-amdhsa -mcpu=gfx900 -verify-machineinstrs | FileCheck -check-prefixes=V5 %s
   4
   5 ; CHECK-LABEL: {{^}}recursive:
   6 ; CHECK: .set recursive.private_seg_size, 16+(max(16384))
   7 ; CHECK: ScratchSize: 16
   8
   9 ; V5-LABEL: {{^}}recursive:
  10 ; V5: .set recursive.has_recursion, 1
  11 define void @recursive() {
  12   call void @recursive()
  13   store volatile i32 0, ptr addrspace(1) undef
  14   ret void
  15 }
  16
  17 ; CHECK-LABEL: {{^}}tail_recursive:
  18 ; CHECK: .set tail_recursive.private_seg_size, 0
  19 ; CHECK: ScratchSize: 0
  20 define void @tail_recursive() {
  21   tail call void @tail_recursive()
  22   ret void
  23 }
  24
  25 ; CHECK: .set calls_tail_recursive.private_seg_size, 0+(max(tail_recursive.private_seg_size))
  26 define void @calls_tail_recursive() norecurse {
  27   tail call void @tail_recursive()
  28   ret void
  29 }
  30
  31 ; CHECK-LABEL: {{^}}tail_recursive_with_stack:
  32 ; CHECK: .set tail_recursive_with_stack.private_seg_size, 8
  33 ; CHECK: .set tail_recursive_with_stack.has_recursion, 1
  34 define void @tail_recursive_with_stack() {
  35   %alloca = alloca i32, addrspace(5)
  36   store volatile i32 0, ptr addrspace(5) %alloca
  37   tail call void @tail_recursive_with_stack()
  38   ret void
  39 }
  40
  41 ; For an arbitrary recursive call, report a large number for unknown stack
  42 ; usage for code object v4 and older
  43 ; CHECK-LABEL: {{^}}calls_recursive:
  44 ; CHECK: .set calls_recursive.private_seg_size, 0+(max(16384, recursive.private_seg_size))
  45 ;
  46 ; V5-LABEL: {{^}}calls_recursive:
  47 ; V5: .set calls_recursive.private_seg_size, 0+(max(recursive.private_seg_size))
  48 ; V5: .set calls_recursive.has_dyn_sized_stack, or(0, recursive.has_dyn_sized_stack)
  49 define amdgpu_kernel void @calls_recursive() {
  50   call void @recursive()
  51   ret void
  52 }
  53
  54 ; Make sure we do not report a huge stack size for tail recursive
  55 ; functions
  56 ; CHECK-LABEL: {{^}}kernel_indirectly_calls_tail_recursive:
  57 ; CHECK: .set kernel_indirectly_calls_tail_recursive.private_seg_size, 0+(max(calls_tail_recursive.private_seg_size))
  58 define amdgpu_kernel void @kernel_indirectly_calls_tail_recursive() {
  59   call void @calls_tail_recursive()
  60   ret void
  61 }
  62
  63 ; TODO: Even though tail_recursive is only called as a tail call, we
  64 ; end up treating it as generally recursive call from the regular call
  65 ; in the kernel.
  66
  67 ; CHECK-LABEL: {{^}}kernel_calls_tail_recursive:
  68 ; CHECK: .set kernel_calls_tail_recursive.private_seg_size, 0+(max(16384, tail_recursive.private_seg_size))
  69 ;
  70 ; V5-LABEL: {{^}}kernel_calls_tail_recursive:
  71 ; V5: .set kernel_calls_tail_recursive.private_seg_size, 0+(max(tail_recursive.private_seg_size))
  72 ; V5: .set kernel_calls_tail_recursive.has_recursion, or(1, tail_recursive.has_recursion)
  73 define amdgpu_kernel void @kernel_calls_tail_recursive() {
  74   call void @tail_recursive()
  75   ret void
  76 }
  77
  78 ; CHECK-LABEL: {{^}}kernel_calls_tail_recursive_with_stack:
  79 ; CHECK: .set kernel_calls_tail_recursive_with_stack.private_seg_size, 0+(max(16384, tail_recursive_with_stack.private_seg_size))
  80 ;
  81 ; V5-LABEL: {{^}}kernel_calls_tail_recursive_with_stack:
  82 ; V5: .set kernel_calls_tail_recursive_with_stack.private_seg_size, 0+(max(tail_recursive_with_stack.private_seg_size))
  83 ; V5: .set kernel_calls_tail_recursive_with_stack.has_dyn_sized_stack, or(0, tail_recursive_with_stack.has_dyn_sized_stack)
  84 define amdgpu_kernel void @kernel_calls_tail_recursive_with_stack() {
  85   call void @tail_recursive_with_stack()
  86   ret void
  87 }
  88
  89 !llvm.module.flags = !{!0}
  90 !0 = !{i32 1, !"amdhsa_code_object_version", i32 CODE_OBJECT_VERSION}