36 lines · plain
1; RUN: llc -mtriple=amdgcn -mcpu=gfx900 < %s | FileCheck -enable-var-scope %s2 3; creating v4i16->v4f16 and v4f16->v4i16 bitcasts in the selection DAG is rather4; difficult, so this test has to throw in some llvm.amdgcn.wqm to get them5 6; CHECK-LABEL: {{^}}test_to_i16:7; CHECK: s_endpgm8define amdgpu_ps void @test_to_i16(ptr addrspace(8) inreg, <4 x half> inreg) #0 {9 %a_tmp = call <4 x half> @llvm.amdgcn.wqm.v4f16(<4 x half> %1)10 %a_i16_tmp = bitcast <4 x half> %a_tmp to <4 x i16>11 %a_i16 = call <4 x i16> @llvm.amdgcn.wqm.v4i16(<4 x i16> %a_i16_tmp)12 13 %a_i32 = bitcast <4 x i16> %a_i16 to <2 x i32>14 call void @llvm.amdgcn.raw.ptr.buffer.store.v2i32(<2 x i32> %a_i32, ptr addrspace(8) %0, i32 0, i32 0, i32 0)15 ret void16}17 18; CHECK-LABEL: {{^}}test_to_half:19; CHECK: s_endpgm20define amdgpu_ps void @test_to_half(ptr addrspace(8) inreg, <4 x i16> inreg) #0 {21 %a_tmp = call <4 x i16> @llvm.amdgcn.wqm.v4i16(<4 x i16> %1)22 %a_half_tmp = bitcast <4 x i16> %a_tmp to <4 x half>23 %a_half = call <4 x half> @llvm.amdgcn.wqm.v4f16(<4 x half> %a_half_tmp)24 25 %a_i32 = bitcast <4 x half> %a_half to <2 x i32>26 call void @llvm.amdgcn.raw.ptr.buffer.store.v2i32(<2 x i32> %a_i32, ptr addrspace(8) %0, i32 0, i32 0, i32 0)27 ret void28}29 30declare <4 x half> @llvm.amdgcn.wqm.v4f16(<4 x half>) #131declare <4 x i16> @llvm.amdgcn.wqm.v4i16(<4 x i16>) #132declare void @llvm.amdgcn.raw.ptr.buffer.store.v2i32(<2 x i32>, ptr addrspace(8), i32, i32, i32) #033 34attributes #0 = { nounwind memory(argmem: write) }35attributes #1 = { nounwind readonly }36