AMDGPU: Fix splitting x16 SMRD loads

When used recursively, this would set the kill flag
on the intermediate step from first splitting
x16 to x8.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@248741 91177308-0d34-0410-b5e6-96231b3b80d8

diff --git a/lib/Target/AMDGPU/SIInstrInfo.cpp b/lib/Target/AMDGPU/SIInstrInfo.cpp
index 432f763d9cba774ccd92777ff73e9595188b2584..5d0cfb32e4cdc719ad317ab3d841249bc78b9faa 100644 (file)
--- a/lib/Target/AMDGPU/SIInstrInfo.cpp
+++ b/lib/Target/AMDGPU/SIInstrInfo.cpp
@@ -2037,8 +2037,8 @@ void SIInstrInfo::splitSMRD(MachineInstr *MI,
                   .addOperand(*SOff);
     unsigned OffsetSGPR = MRI.createVirtualRegister(&AMDGPU::SReg_32RegClass);
     BuildMI(*MBB, MI, DL, get(AMDGPU::S_ADD_I32), OffsetSGPR)
-            .addOperand(*SOff)
-            .addImm(HalfSize);
+      .addReg(SOff->getReg(), 0, SOff->getSubReg())
+      .addImm(HalfSize);
     Hi = BuildMI(*MBB, MI, DL, get(HalfSGPROp), RegHi)
                   .addReg(SBase->getReg(), getKillRegState(IsKill),
                           SBase->getSubReg())

diff --git a/test/CodeGen/AMDGPU/salu-to-valu.ll b/test/CodeGen/AMDGPU/salu-to-valu.ll
index de798182271cd68bf8c64ea944e68df9b5c68e50..65b4a7dadba952b833e7cd63b3ee9ff1f748702e 100644 (file)
--- a/test/CodeGen/AMDGPU/salu-to-valu.ll
+++ b/test/CodeGen/AMDGPU/salu-to-valu.ll
@@ -181,6 +181,49 @@ entry:
   ret void
 }
 
+; FIXME: should use immediate offset instead of using s_add_i32 for adding to constant.
+; GCN-LABEL: {{^}}smrd_valu_ci_offset_x16:
+
+; GCN: s_mov_b32 s[[OFFSET0:[0-9]+]], 0x13480{{$}}
+; SI: s_add_i32 s[[OFFSET1:[0-9]+]], s[[OFFSET0]], 16
+; GCN: buffer_load_dwordx4 v{{\[[0-9]+:[0-9]+\]}}, v{{\[[0-9]+:[0-9]+\]}}, s{{\[}}[[OFFSET0]]:{{[0-9]+}}], 0 addr64{{$}}
+
+; CI: s_mov_b32 s[[OFFSET1:[0-9]+]], 0x13490{{$}}
+; GCN: buffer_load_dwordx4 v{{\[[0-9]+:[0-9]+\]}}, v{{\[[0-9]+:[0-9]+\]}}, s{{\[}}[[OFFSET1]]:{{[0-9]+}}], 0 addr64{{$}}
+
+; SI: s_add_i32 s[[OFFSET2:[0-9]+]], s[[OFFSET0]], 32
+; CI: s_mov_b32 s[[OFFSET2:[0-9]+]], 0x134a0
+
+; GCN: buffer_load_dwordx4 v{{\[[0-9]+:[0-9]+\]}}, v{{\[[0-9]+:[0-9]+\]}}, s{{\[}}[[OFFSET2]]:{{[0-9]+}}], 0 addr64{{$}}
+; GCN: s_add_i32 s[[OFFSET3:[0-9]+]], s[[OFFSET2]], 16
+; GCN: buffer_load_dwordx4 v{{\[[0-9]+:[0-9]+\]}}, v{{\[[0-9]+:[0-9]+\]}}, s{{\[}}[[OFFSET3]]:{{[0-9]+}}], 0 addr64{{$}}
+; GCN: v_or_b32_e32 {{v[0-9]+}}, {{s[0-9]+}}, {{v[0-9]+}}
+; GCN: v_or_b32_e32 {{v[0-9]+}}, {{s[0-9]+}}, {{v[0-9]+}}
+; GCN: v_or_b32_e32 {{v[0-9]+}}, {{s[0-9]+}}, {{v[0-9]+}}
+; GCN: v_or_b32_e32 {{v[0-9]+}}, {{s[0-9]+}}, {{v[0-9]+}}
+; GCN: v_or_b32_e32 {{v[0-9]+}}, {{s[0-9]+}}, {{v[0-9]+}}
+; GCN: v_or_b32_e32 {{v[0-9]+}}, {{s[0-9]+}}, {{v[0-9]+}}
+; GCN: v_or_b32_e32 {{v[0-9]+}}, {{s[0-9]+}}, {{v[0-9]+}}
+; GCN: v_or_b32_e32 {{v[0-9]+}}, {{s[0-9]+}}, {{v[0-9]+}}
+; GCN: buffer_store_dword
+; GCN: buffer_store_dword
+; GCN: buffer_store_dword
+; GCN: buffer_store_dword
+; GCN: buffer_store_dword
+; GCN: buffer_store_dword
+; GCN: buffer_store_dword
+; GCN: buffer_store_dword
+define void @smrd_valu_ci_offset_x16(<16 x i32> addrspace(1)* %out, <16 x i32> addrspace(2)* %in, <16 x i32> %c) #1 {
+entry:
+  %tmp = call i32 @llvm.r600.read.tidig.x() #0
+  %tmp2 = getelementptr <16 x i32>, <16 x i32> addrspace(2)* %in, i32 %tmp
+  %tmp3 = getelementptr <16 x i32>, <16 x i32> addrspace(2)* %tmp2, i32 1234
+  %tmp4 = load <16 x i32>, <16 x i32> addrspace(2)* %tmp3
+  %tmp5 = or <16 x i32> %tmp4, %c
+  store <16 x i32> %tmp5, <16 x i32> addrspace(1)* %out
+  ret void
+}
+
 ; GCN-LABEL: {{^}}smrd_valu2_salu_user:
 ; GCN: buffer_load_dword [[MOVED:v[0-9]+]], v{{\[[0-9]+:[0-9]+\]}}, s{{\[[0-9]+:[0-9]+\]}}, 0 addr64 offset:16{{$}}
 ; GCN: v_add_i32_e32 [[ADD:v[0-9]+]], vcc, s{{[0-9]+}}, [[MOVED]]