• Home
  • Line#
  • Scopes#
  • Navigate#
  • Raw
  • Download
1; RUN: llc -march=amdgcn -verify-machineinstrs < %s | FileCheck -check-prefix=SI %s
2
3
4declare i32 @llvm.amdgcn.workitem.id.x() readnone
5
6; SI-LABEL: {{^}}test_i64_vreg:
7; SI: v_add_i32
8; SI: v_addc_u32
9define amdgpu_kernel void @test_i64_vreg(i64 addrspace(1)* noalias %out, i64 addrspace(1)* noalias %inA, i64 addrspace(1)* noalias %inB) {
10  %tid = call i32 @llvm.amdgcn.workitem.id.x() readnone
11  %a_ptr = getelementptr i64, i64 addrspace(1)* %inA, i32 %tid
12  %b_ptr = getelementptr i64, i64 addrspace(1)* %inB, i32 %tid
13  %a = load i64, i64 addrspace(1)* %a_ptr
14  %b = load i64, i64 addrspace(1)* %b_ptr
15  %result = add i64 %a, %b
16  store i64 %result, i64 addrspace(1)* %out
17  ret void
18}
19
20; Check that the SGPR add operand is correctly moved to a VGPR.
21; SI-LABEL: {{^}}sgpr_operand:
22; SI: s_add_u32
23; SI: s_addc_u32
24define amdgpu_kernel void @sgpr_operand(i64 addrspace(1)* noalias %out, i64 addrspace(1)* noalias %in, i64 addrspace(1)* noalias %in_bar, i64 %a) {
25  %foo = load i64, i64 addrspace(1)* %in, align 8
26  %result = add i64 %foo, %a
27  store i64 %result, i64 addrspace(1)* %out
28  ret void
29}
30
31; Swap the arguments. Check that the SGPR -> VGPR copy works with the
32; SGPR as other operand.
33;
34; SI-LABEL: {{^}}sgpr_operand_reversed:
35; SI: s_add_u32
36; SI: s_addc_u32
37define amdgpu_kernel void @sgpr_operand_reversed(i64 addrspace(1)* noalias %out, i64 addrspace(1)* noalias %in, i64 %a) {
38  %foo = load i64, i64 addrspace(1)* %in, align 8
39  %result = add i64 %a, %foo
40  store i64 %result, i64 addrspace(1)* %out
41  ret void
42}
43
44
45; SI-LABEL: {{^}}test_v2i64_sreg:
46; SI: s_add_u32
47; SI: s_addc_u32
48; SI: s_add_u32
49; SI: s_addc_u32
50define amdgpu_kernel void @test_v2i64_sreg(<2 x i64> addrspace(1)* noalias %out, <2 x i64> %a, <2 x i64> %b) {
51  %result = add <2 x i64> %a, %b
52  store <2 x i64> %result, <2 x i64> addrspace(1)* %out
53  ret void
54}
55
56; SI-LABEL: {{^}}test_v2i64_vreg:
57; SI: v_add_i32
58; SI: v_addc_u32
59; SI: v_add_i32
60; SI: v_addc_u32
61define amdgpu_kernel void @test_v2i64_vreg(<2 x i64> addrspace(1)* noalias %out, <2 x i64> addrspace(1)* noalias %inA, <2 x i64> addrspace(1)* noalias %inB) {
62  %tid = call i32 @llvm.amdgcn.workitem.id.x() readnone
63  %a_ptr = getelementptr <2 x i64>, <2 x i64> addrspace(1)* %inA, i32 %tid
64  %b_ptr = getelementptr <2 x i64>, <2 x i64> addrspace(1)* %inB, i32 %tid
65  %a = load <2 x i64>, <2 x i64> addrspace(1)* %a_ptr
66  %b = load <2 x i64>, <2 x i64> addrspace(1)* %b_ptr
67  %result = add <2 x i64> %a, %b
68  store <2 x i64> %result, <2 x i64> addrspace(1)* %out
69  ret void
70}
71
72; SI-LABEL: {{^}}trunc_i64_add_to_i32:
73; SI: s_load_dword s[[SREG0:[0-9]+]]
74; SI: s_load_dword s[[SREG1:[0-9]+]]
75; SI: s_add_i32 [[SRESULT:s[0-9]+]], s[[SREG1]], s[[SREG0]]
76; SI-NOT: addc
77; SI: v_mov_b32_e32 [[VRESULT:v[0-9]+]], [[SRESULT]]
78; SI: buffer_store_dword [[VRESULT]],
79define amdgpu_kernel void @trunc_i64_add_to_i32(i32 addrspace(1)* %out, i32, i64 %a, i32, i64 %b) {
80  %add = add i64 %b, %a
81  %trunc = trunc i64 %add to i32
82  store i32 %trunc, i32 addrspace(1)* %out, align 8
83  ret void
84}
85