1 ; NOTE: Assertions have been autogenerated by utils/update_llc_test_checks.py
2 ; RUN: llc -mtriple=amdgcn-amd-amdhsa -mcpu=gfx908 -verify-machineinstrs -o - %s | FileCheck -check-prefix=GCN %s
3 ; RUN: llc -mtriple=amdgcn-amd-amdhsa -mcpu=gfx90a -verify-machineinstrs -o - %s | FileCheck -check-prefix=GCN %s
5 ; Callee must preserve the VGPR modified by writelane even if it is marked Caller-saved.
7 declare i32 @llvm.amdgcn.writelane(i32, i32, i32)
9 define void @sgpr_spill_writelane() {
10 ; GCN-LABEL: sgpr_spill_writelane:
12 ; GCN-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
13 ; GCN-NEXT: s_xor_saveexec_b64 s[4:5], -1
14 ; GCN-NEXT: buffer_store_dword v0, off, s[0:3], s32 ; 4-byte Folded Spill
15 ; GCN-NEXT: s_mov_b64 exec, s[4:5]
16 ; GCN-NEXT: v_writelane_b32 v0, s35, 0
17 ; GCN-NEXT: ;;#ASMSTART
19 ; GCN-NEXT: v_readlane_b32 s35, v0, 0
20 ; GCN-NEXT: s_xor_saveexec_b64 s[4:5], -1
21 ; GCN-NEXT: buffer_load_dword v0, off, s[0:3], s32 ; 4-byte Folded Reload
22 ; GCN-NEXT: s_mov_b64 exec, s[4:5]
23 ; GCN-NEXT: s_waitcnt vmcnt(0)
24 ; GCN-NEXT: s_setpc_b64 s[30:31]
25 call void asm sideeffect "", "~{s35}"()
29 ; FIXME: The writelane intrinsic doesn't really overwrite any inactive lanes
30 ; and hence there is no need to preserve the VGPR it modifies.
31 define void @device_writelane_intrinsic(ptr addrspace(1) %out, i32 %src) {
32 ; GCN-LABEL: device_writelane_intrinsic:
34 ; GCN-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
35 ; GCN-NEXT: s_xor_saveexec_b64 s[4:5], -1
36 ; GCN-NEXT: buffer_store_dword v3, off, s[0:3], s32 ; 4-byte Folded Spill
37 ; GCN-NEXT: s_mov_b64 exec, s[4:5]
38 ; GCN-NEXT: v_mov_b32_e32 v3, 15
39 ; GCN-NEXT: v_readfirstlane_b32 s4, v2
40 ; GCN-NEXT: v_writelane_b32 v3, s4, 23
41 ; GCN-NEXT: global_store_dword v[0:1], v3, off
42 ; GCN-NEXT: s_xor_saveexec_b64 s[4:5], -1
43 ; GCN-NEXT: buffer_load_dword v3, off, s[0:3], s32 ; 4-byte Folded Reload
44 ; GCN-NEXT: s_mov_b64 exec, s[4:5]
45 ; GCN-NEXT: s_waitcnt vmcnt(0)
46 ; GCN-NEXT: s_setpc_b64 s[30:31]
47 %writelane = call i32 @llvm.amdgcn.writelane(i32 %src, i32 23, i32 15)
48 store i32 %writelane, ptr addrspace(1) %out, align 4
52 define amdgpu_kernel void @kernel_writelane_intrinsic(ptr addrspace(1) %out, i32 %src0, i32 %src1) {
53 ; GCN-LABEL: kernel_writelane_intrinsic:
55 ; GCN-NEXT: s_load_dwordx4 s[0:3], s[4:5], 0x0
56 ; GCN-NEXT: v_mov_b32_e32 v1, 45
57 ; GCN-NEXT: v_mov_b32_e32 v0, 0
58 ; GCN-NEXT: s_waitcnt lgkmcnt(0)
59 ; GCN-NEXT: s_mov_b32 m0, s3
60 ; GCN-NEXT: v_writelane_b32 v1, s2, m0
61 ; GCN-NEXT: global_store_dword v0, v1, s[0:1]
63 %writelane = call i32 @llvm.amdgcn.writelane(i32 %src0, i32 %src1, i32 45)
64 store i32 %writelane, ptr addrspace(1) %out, align 4