AMDGPU: fix overlapping copies in copyPhysReg

[nhaehnle]

Summary

When copying aggregate registers within the same register class, there may
be an overlap between source and destination that forces us to do the copy
backwards.

Do the simplest possible thing (relying on the order of enums generated by
tablegen) that guarantees the correct order of moves when there are overlaps,
and does whatever when there is no overlap. (The last part forces some
trivial adjustments to test cases.)

Together with r255906, this fixes a VM fault in Unreal Elemental Demo.

While at it, change the generation of kill and def flags to something that
looks more reasonable. This method is used very late during compilation, so
it probably doesn't matter in practice, and to be honest, I don't know if
this change is actually correct because the semantics in connection with
aggregate registers vs. sub-registers are not clear to me.

Bugzilla: https://bugs.freedesktop.org/show_bug.cgi?id=93264

[• tstellarAMD]

lib/Target/AMDGPU/SIInstrInfo.cpp

433

You should use SIRegisterInfo::getHWRegIndex(unsigned Reg) here so we don't have to rely on the order of the enumerations.

[nhaehnle]

That makes more sense indeed.

Had to adjust two tests yet again, so I decided to convert the flat-address-space
one so that it uses -DAG. For some reason, -DAG between -NOTs didn't work properly
for the move-addr64-rsrc-dead-subreg-writes one, so I just switched the lines
around again.

[• tstellarAMD]

lib/Target/AMDGPU/SIInstrInfo.cpp

431–438

Is there some reason why you used a variable here and didn't put this condition directly in the if() below?

[nhaehnle]

lib/Target/AMDGPU/SIInstrInfo.cpp

431–438

Yes: the intention was to make it obvious that the condition is constant throughout the loop for both the human and the compiler.

[• tstellarAMD]

LGTM.

lib/Target/AMDGPU/SIInstrInfo.cpp

SIInstrInfo::copyPhysReg(MachineBasicBlock &MBB,
   // the backend.  While it may be theoretically possible to do this, it should
   // never be necessary.
   assert(DestReg != AMDGPU::SCC && SrcReg != AMDGPU::SCC);
 
   static const int16_t Sub0_15[] = {
     AMDGPU::sub0, AMDGPU::sub1, AMDGPU::sub2, AMDGPU::sub3,
     AMDGPU::sub4, AMDGPU::sub5, AMDGPU::sub6, AMDGPU::sub7,
     AMDGPU::sub8, AMDGPU::sub9, AMDGPU::sub10, AMDGPU::sub11,
-    AMDGPU::sub12, AMDGPU::sub13, AMDGPU::sub14, AMDGPU::sub15, 0
+    AMDGPU::sub12, AMDGPU::sub13, AMDGPU::sub14, AMDGPU::sub15,
   };
 
   static const int16_t Sub0_7[] = {
     AMDGPU::sub0, AMDGPU::sub1, AMDGPU::sub2, AMDGPU::sub3,
-    AMDGPU::sub4, AMDGPU::sub5, AMDGPU::sub6, AMDGPU::sub7, 0
+    AMDGPU::sub4, AMDGPU::sub5, AMDGPU::sub6, AMDGPU::sub7,
   };
 
   static const int16_t Sub0_3[] = {
-    AMDGPU::sub0, AMDGPU::sub1, AMDGPU::sub2, AMDGPU::sub3, 0
+    AMDGPU::sub0, AMDGPU::sub1, AMDGPU::sub2, AMDGPU::sub3,
   };
 
   static const int16_t Sub0_2[] = {
-    AMDGPU::sub0, AMDGPU::sub1, AMDGPU::sub2, 0
+    AMDGPU::sub0, AMDGPU::sub1, AMDGPU::sub2,
   };
 
   static const int16_t Sub0_1[] = {
-    AMDGPU::sub0, AMDGPU::sub1, 0
+    AMDGPU::sub0, AMDGPU::sub1,
   };
 
   unsigned Opcode;
-  const int16_t *SubIndices;
+  ArrayRef<int16_t> SubIndices;
 
   if (AMDGPU::SReg_32RegClass.contains(DestReg)) {
     assert(AMDGPU::SReg_32RegClass.contains(SrcReg));
     BuildMI(MBB, MI, DL, get(AMDGPU::S_MOV_B32), DestReg)
             .addReg(SrcReg, getKillRegState(KillSrc));
     return;
 
   } else if (AMDGPU::SReg_64RegClass.contains(DestReg)) {
assert(AMDGPU::VReg_512RegClass.contains(SrcReg) ||
            AMDGPU::SReg_512RegClass.contains(SrcReg));
     Opcode = AMDGPU::V_MOV_B32_e32;
     SubIndices = Sub0_15;
 
   } else {
     llvm_unreachable("Can't copy register!");
   }
 
-  while (unsigned SubIdx = *SubIndices++) {
+  for (unsigned Idx = 0; Idx < SubIndices.size(); ++Idx) {
+    unsigned SubIdx;
+    if (DestReg <= SrcReg)
+      SubIdx = SubIndices[Idx];
+    else
+      SubIdx = SubIndices[SubIndices.size() - Idx - 1];
+
     MachineInstrBuilder Builder = BuildMI(MBB, MI, DL,
       get(Opcode), RI.getSubReg(DestReg, SubIdx));
 
-    Builder.addReg(RI.getSubReg(SrcReg, SubIdx), getKillRegState(KillSrc));
+    Builder.addReg(RI.getSubReg(SrcReg, SubIdx));
+
+    if (Idx == SubIndices.size() - 1)
+      Builder.addReg(SrcReg, RegState::Kill | RegState::Implicit);
 
-    if (*SubIndices)
+    if (Idx == 0)
       Builder.addReg(DestReg, RegState::Define | RegState::Implicit);
   }
 }
 
 int SIInstrInfo::commuteOpcode(const MachineInstr &MI) const {
   const unsigned Opcode = MI.getOpcode();
 
   int NewOpc;

test/CodeGen/AMDGPU/ctpop64.ll

 
 ; FIXME: We currently disallow SALU instructions in all branches,
 ; but there are some cases when the should be allowed.
 
 ; FUNC-LABEL: {{^}}ctpop_i64_in_br:
 ; SI: s_load_dwordx2 s{{\[}}[[LOVAL:[0-9]+]]:[[HIVAL:[0-9]+]]{{\]}}, s[{{[0-9]+:[0-9]+}}], 0xd
 ; VI: s_load_dwordx2 s{{\[}}[[LOVAL:[0-9]+]]:[[HIVAL:[0-9]+]]{{\]}}, s[{{[0-9]+:[0-9]+}}], 0x34
 ; GCN: s_bcnt1_i32_b64 [[RESULT:s[0-9]+]], {{s\[}}[[LOVAL]]:[[HIVAL]]{{\]}}
-; GCN: v_mov_b32_e32 v[[VLO:[0-9]+]], [[RESULT]]
-; GCN: v_mov_b32_e32 v[[VHI:[0-9]+]], s[[HIVAL]]
+; GCN-DAG: v_mov_b32_e32 v[[VLO:[0-9]+]], [[RESULT]]
+; GCN-DAG: v_mov_b32_e32 v[[VHI:[0-9]+]], s[[HIVAL]]
 ; GCN: buffer_store_dwordx2 {{v\[}}[[VLO]]:[[VHI]]{{\]}}
 ; GCN: s_endpgm
 define void @ctpop_i64_in_br(i64 addrspace(1)* %out, i64 addrspace(1)* %in, i64 %ctpop_arg, i32 %cond) {
 entry:
   %tmp0 = icmp eq i32 %cond, 0
   br i1 %tmp0, label %if, label %else
 
 if:

test/CodeGen/AMDGPU/flat-address-space.ll

 ; RUN: llc -O0 -march=amdgcn -mcpu=bonaire -mattr=-promote-alloca < %s | FileCheck -check-prefix=CHECK -check-prefix=CHECK-NO-PROMOTE %s
 ; RUN: llc -O0 -march=amdgcn -mcpu=bonaire -mattr=+promote-alloca < %s | FileCheck -check-prefix=CHECK -check-prefix=CHECK-PROMOTE %s
 ; RUN: llc -O0 -march=amdgcn -mcpu=tonga -mattr=-promote-alloca < %s | FileCheck -check-prefix=CHECK -check-prefix=CHECK-NO-PROMOTE %s
 ; RUN: llc -O0 -march=amdgcn -mcpu=tonga -mattr=+promote-alloca < %s | FileCheck -check-prefix=CHECK -check-prefix=CHECK-PROMOTE %s
 
 ; Disable optimizations in case there are optimizations added that
 ; specialize away generic pointer accesses.
 
 
 ; These testcases might become useless when there are optimizations to
 ; remove generic pointers.
 
 ; CHECK-LABEL: {{^}}store_flat_i32:
 ; CHECK: v_mov_b32_e32 v[[DATA:[0-9]+]], {{s[0-9]+}}
-; CHECK: v_mov_b32_e32 v[[LO_VREG:[0-9]+]], {{s[0-9]+}}
 ; CHECK: v_mov_b32_e32 v[[HI_VREG:[0-9]+]], {{s[0-9]+}}
+; CHECK: v_mov_b32_e32 v[[LO_VREG:[0-9]+]], {{s[0-9]+}}
 ; CHECK: flat_store_dword v[[DATA]], v{{\[}}[[LO_VREG]]:[[HI_VREG]]{{\]}}
 define void @store_flat_i32(i32 addrspace(1)* %gptr, i32 %x) #0 {
   %fptr = addrspacecast i32 addrspace(1)* %gptr to i32 addrspace(4)*
   store i32 %x, i32 addrspace(4)* %fptr, align 4
   ret void
 }
 
 ; CHECK-LABEL: {{^}}store_flat_i64:

test/CodeGen/AMDGPU/move-addr64-rsrc-dead-subreg-writes.ll

 ; RUN: llc -march=amdgcn -mcpu=kaveri -mtriple=amdgcn-unknown-amdhsa < %s | FileCheck -check-prefix=GCN %s
 
 ; Check that when mubuf addr64 instruction is handled in moveToVALU
 ; from the pointer, dead register writes are not emitted.
 
 ; FIXME: We should be able to use the SGPR directly as src0 to v_add_i32
 
 ; GCN-LABEL: {{^}}clobber_vgpr_pair_pointer_add:
 ; GCN: s_load_dwordx2 s{{\[}}[[ARG1LO:[0-9]+]]:[[ARG1HI:[0-9]+]]{{\]}}, s{{\[[0-9]+:[0-9]+\]}}, 0x0{{$}}
 ; GCN: buffer_load_dwordx2 v{{\[}}[[LDPTRLO:[0-9]+]]:[[LDPTRHI:[0-9]+]]{{\]}}
 
 ; GCN-NOT: v_mov_b32
-; GCN: v_mov_b32_e32 v[[VARG1LO:[0-9]+]], s[[ARG1LO]]
-; GCN-NEXT: v_mov_b32_e32 v[[VARG1HI:[0-9]+]], s[[ARG1HI]]
+; GCN: v_mov_b32_e32 v[[VARG1HI:[0-9]+]], s[[ARG1HI]]
+; GCN-NEXT: v_mov_b32_e32 v[[VARG1LO:[0-9]+]], s[[ARG1LO]]
 ; GCN-NOT: v_mov_b32
 
 ; GCN: v_add_i32_e32 v[[PTRLO:[0-9]+]], vcc, v[[LDPTRLO]], v[[VARG1LO]]
 ; GCN: v_addc_u32_e32 v[[PTRHI:[0-9]+]], vcc, v[[LDPTRHI]], v[[VARG1HI]]
 ; GCN: buffer_load_ubyte v{{[0-9]+}}, v{{\[}}[[PTRLO]]:[[PTRHI]]{{\]}},
 
 define void @clobber_vgpr_pair_pointer_add(i64 %arg1, i8 addrspace(1)* addrspace(1)* %ptrarg, i32 %arg3) #0 {
 bb: