xref: /llvm-project/llvm/test/CodeGen/AMDGPU/selectcc-opt.ll (revision 9e9907f1cfa424366fba58d9520f9305b537cec9)
1; RUN: llc -mtriple=amdgcn < %s | FileCheck -check-prefix=SI -check-prefix=FUNC %s
2; RUN: llc -mtriple=amdgcn -mcpu=tonga -mattr=-flat-for-global < %s | FileCheck -check-prefix=SI -check-prefix=FUNC %s
3; RUN: llc -mtriple=r600 -mcpu=redwood < %s | FileCheck -check-prefix=EG -check-prefix=FUNC %s
4
5
6; FUNC-LABEL: {{^}}test_a:
7; EG-NOT: CND
8; EG: SET{{[NEQGTL]+}}_DX10
9
10define amdgpu_kernel void @test_a(ptr addrspace(1) %out, float %in) {
11entry:
12  %0 = fcmp olt float %in, 0.000000e+00
13  %1 = select i1 %0, float 1.000000e+00, float 0.000000e+00
14  %2 = fneg float %1
15  %3 = fptosi float %2 to i32
16  %4 = bitcast i32 %3 to float
17  %5 = bitcast float %4 to i32
18  %6 = icmp ne i32 %5, 0
19  br i1 %6, label %IF, label %ENDIF
20
21IF:
22  %7 = getelementptr i32, ptr addrspace(1) %out, i32 1
23  store i32 0, ptr addrspace(1) %7
24  br label %ENDIF
25
26ENDIF:
27  store i32 0, ptr addrspace(1) %out
28  ret void
29}
30
31; Same as test_a, but the branch labels are swapped to produce the inverse cc
32; for the icmp instruction
33
34; EG-LABEL: {{^}}test_b:
35; EG: SET{{[GTEQN]+}}_DX10
36; EG-NEXT: 0(0.000000e+00), 0(0.000000e+00)
37; EG-NEXT: PRED_
38; EG-NEXT: ALU clause starting
39define amdgpu_kernel void @test_b(ptr addrspace(1) %out, float %in) {
40entry:
41  %0 = fcmp olt float %in, 0.0
42  %1 = select i1 %0, float 1.000000e+00, float 0.000000e+00
43  %2 = fneg float %1
44  %3 = fptosi float %2 to i32
45  %4 = bitcast i32 %3 to float
46  %5 = bitcast float %4 to i32
47  %6 = icmp ne i32 %5, 0
48  br i1 %6, label %ENDIF, label %IF
49
50IF:
51  %7 = getelementptr i32, ptr addrspace(1) %out, i32 1
52  store i32 0, ptr addrspace(1) %7
53  br label %ENDIF
54
55ENDIF:
56  store i32 0, ptr addrspace(1) %out
57  ret void
58}
59
60; Test a CND*_INT instruction with float true/false values
61; EG-LABEL: {{^}}test_c:
62; EG: CND{{[GTE]+}}_INT
63define amdgpu_kernel void @test_c(ptr addrspace(1) %out, i32 %in) {
64entry:
65  %0 = icmp sgt i32 %in, 0
66  %1 = select i1 %0, float 2.0, float 3.0
67  store float %1, ptr addrspace(1) %out
68  ret void
69}
70
71; FUNC-LABEL: {{^}}selectcc_bool:
72; SI: s_cmp_lg_u32
73; SI: v_cndmask_b32_e64
74; SI-NOT: cmp
75; SI-NOT: cndmask
76define amdgpu_kernel void @selectcc_bool(ptr addrspace(1) %out, i32 %a, i32 %b) nounwind {
77  %icmp0 = icmp ne i32 %a, %b
78  %ext = select i1 %icmp0, i32 -1, i32 0
79  store i32 %ext, ptr addrspace(1) %out
80  ret void
81}
82