[AMDGPU] Add peephole to optimize MOV

[piotr]

Summary

Add peephole optimization to remove redundant MOV if:

• the MOV is identical to a MOV in the immediate predecessor of MBB and
• no instruction between them modifies the destination register

Change-Id: Ibb187e3219ef641b7681e4123c7ca3fe8e1d14e6

[arsenm]

This doesn't seem like the right place to handle this. This doesn't really fit with SIShrinkInstructions. I also wouldn't expect you need to do anything special to handle this case. Why doesn't MachineCSE catch this? Can you add an IR example testcase?

[rampitec]

Patch does not take control flow divergence into account.

lib/Target/AMDGPU/SIShrinkInstructions.cpp

682

Even is block has only one predecessor that does not mean this is the reaching def of the physreg. The predecessor may (and likely will) have a different exec mask. It might work with sclaral, but not with vector moves.

[piotr]

This opt opportunity presents itself after the register coalescer, where inline constants are moved to the same registers (attaching IR example testcase imminently). MachineCSE is run too early to spot this and the last occurence of si-shrink-instructions is run at the right time. I think, for similar reasons some other peepholes were also placed in this pass, even though they are not about 32-bit encoding.

[piotr]

lib/Target/AMDGPU/SIShrinkInstructions.cpp

682

Perhaps I misunderstood your comment, but if the block has only one predecessor then it is executed for a subset of threads that already executed the predecessor, so for that subset we're sure that the previous mov happened?

[piotr]

Adding IR test case example.

[rampitec]

lib/Target/AMDGPU/SIShrinkInstructions.cpp

682

Consider this:

a:
  v_mov_b32 v0, 1
  s_and_savexec_b64 s[0:1], vcc
  ; mask branch %c
b:
  v_mov_b32 v0, 2
c:
  s_or_b32 exec, exec, s[0:1]
  use v0

You will have divergent v0 value at use even though block c has only one predecessor.

[arsenm]

In D63709#1555766, @piotr wrote:

This opt opportunity presents itself after the register coalescer, where inline constants are moved to the same registers (attaching IR example testcase imminently). MachineCSE is run too early to spot this and the last occurence of si-shrink-instructions is run at the right time. I think, for similar reasons some other peepholes were also placed in this pass, even though they are not about 32-bit encoding.

I'd like to take a look a more reduced testcase. My intuition is something else is going wrong that's not eliminating this

test/CodeGen/AMDGPU/mov-opt.ll

14 ↗

(On Diff #206233)

You should be able to reduce this more

[arsenm]

I do suspect the handling of the EXEC physreg use in MachineCSE. I don't see any of the typical sources of late visible constants in this testcase?

[arsenm]

test/CodeGen/AMDGPU/mov-opt.ll

14 ↗

(On Diff #206233)

I don't actually see any redundant constants?

[piotr]

test/CodeGen/AMDGPU/mov-opt.ll

14 ↗

(On Diff #206233)

The move constants are by-products of different phi nodes and this is why they are not explicitly written in the IR. They get created during the isel in HandlePHINodesInSuccessorBlocks.

I will work on simplifying the IR test case even more.

[piotr]

Simplifying IR test case even more with bugpoint.

[piotr]

lib/Target/AMDGPU/SIShrinkInstructions.cpp

682

Thanks Stas. Good point, I will need to check for the modification of the exec mask between the movs.

[arsenm]

Reduced further: https://paste.debian.net/1089194

[arsenm]

In D63709#1557592, @arsenm wrote:

Reduced further: https://paste.debian.net/1089194

Other targets seem to not have this problem with a slightly generalized version, so I would look into how this is cleaned up there

[arsenm]

In D63709#1557605, @arsenm wrote:

In D63709#1557592, @arsenm wrote:

Reduced further: https://paste.debian.net/1089194

Other targets seem to not have this problem with a slightly generalized version, so I would look into how this is cleaned up there

It seems we're missing a simplifycfg run somewhere, so maybe we're thinking of this on the wrong level entirely.. If I run simplify cfg on any of the testcase variants, this problem disappears

[arsenm]

In D63709#1557626, @arsenm wrote:

In D63709#1557605, @arsenm wrote:

In D63709#1557592, @arsenm wrote:

Reduced further: https://paste.debian.net/1089194

Other targets seem to not have this problem with a slightly generalized version, so I would look into how this is cleaned up there

It seems we're missing a simplifycfg run somewhere, so maybe we're thinking of this on the wrong level entirely.. If I run simplify cfg on any of the testcase variants, this problem disappears

Other targets seem to run SimiplifyCFG after AtomicExpand, which we are missing. Even with that disabled and the phi survives to machineinstrs, aarch64 and hexagon both avoid this

[piotr]

In D63709#1557626, @arsenm wrote:

In D63709#1557605, @arsenm wrote:

In D63709#1557592, @arsenm wrote:

Reduced further: https://paste.debian.net/1089194

Other targets seem to not have this problem with a slightly generalized version, so I would look into how this is cleaned up there

It seems we're missing a simplifycfg run somewhere, so maybe we're thinking of this on the wrong level entirely.. If I run simplify cfg on any of the testcase variants, this problem disappears

Well spotted, I will try to look at the issue from the perspective of missing simplifycfg.

[arsenm]

In D63709#1557646, @arsenm wrote:

In D63709#1557626, @arsenm wrote:

In D63709#1557605, @arsenm wrote:

In D63709#1557592, @arsenm wrote:

Reduced further: https://paste.debian.net/1089194

Other targets seem to not have this problem with a slightly generalized version, so I would look into how this is cleaned up there

It seems we're missing a simplifycfg run somewhere, so maybe we're thinking of this on the wrong level entirely.. If I run simplify cfg on any of the testcase variants, this problem disappears

Other targets seem to run SimiplifyCFG after AtomicExpand, which we are missing. Even with that disabled and the phi survives to machineinstrs, aarch64 and hexagon both avoid this

MachineCSE should be taking care of this, but it for some reason concludes it isn't profitable:
Examining: %9:vgpr_32 = V_MOV_B32_e32 1065353216, implicit $exec

• Found a common subexpression: %7:vgpr_32 = V_MOV_B32_e32 1065353216, implicit $exec
• Not profitable, avoid CSE!

[arsenm]

I think things are going from from this heuristic:

// Heuristics #3: If the common subexpression is used by PHIs, do not reuse
// it unless the defined value is already used in the BB of the new use.
bool HasPHI = false;
for (MachineInstr &UseMI : MRI->use_nodbg_instructions(CSReg)) {
  HasPHI |= UseMI.isPHI();
  if (UseMI.getParent() == MI->getParent())
    return true;
}

The second phi use is in a flow block with only the phi, so the use is in the fall through successor. Maybe it could be relaxed to look through trivial successors, at least for blocks with only phis?

[piotr]

In D63709#1557689, @arsenm wrote:

I think things are going from from this heuristic:

// Heuristics #3: If the common subexpression is used by PHIs, do not reuse
// it unless the defined value is already used in the BB of the new use.
bool HasPHI = false;
for (MachineInstr &UseMI : MRI->use_nodbg_instructions(CSReg)) {
  HasPHI |= UseMI.isPHI();
  if (UseMI.getParent() == MI->getParent())
    return true;
}

The second phi use is in a flow block with only the phi, so the use is in the fall through successor. Maybe it could be relaxed to look through trivial successors, at least for blocks with only phis?

Nice tracking down. I will abandon this review, and submit another one where I will extend the MachineCSE pass as you suggested.

lib/Target/AMDGPU/SIShrinkInstructions.cpp

if (TargetRegisterInfo::isPhysicalRegister(Reg)) {
       LaneBitmask LM = TRI.getSubRegIndexLaneMask(Sub);
       Sub = TRI.getSubRegFromChannel(I + countTrailingZeros(LM.getAsInteger()));
     }
   }
   return TargetInstrInfo::RegSubRegPair(Reg, Sub);
 }
 
 // Match:
+// bb0:
+// ..
+// mov r, imm
+// ..
+// branch bb1
+//
+// bb1:
+// ; predecessors: %bb.0
+// ..
+// mov r, imm <== redundant mov
+//
+//
+// Returns true if the mov can be removed.
+//
+static bool isMovRedundant(MachineInstr &Mov, MachineRegisterInfo &MRI,
+                           const SIInstrInfo *TII) {
+
+  assert(Mov.getOpcode() == AMDGPU::V_MOV_B32_e32 ||
+         Mov.getOpcode() == AMDGPU::S_MOV_B32 ||
+         Mov.getOpcode() == AMDGPU::S_MOV_B64);
+  assert(Mov.getParent()->pred_size() == 1);
+
+  auto DstOperand = Mov.getOpcode() == AMDGPU::V_MOV_B32_e32
+                        ? TII->getNamedOperand(Mov, AMDGPU::OpName::vdst)
+                        : TII->getNamedOperand(Mov, AMDGPU::OpName::sdst);
+
+  unsigned R = DstOperand->getReg();
+  unsigned Rsub = DstOperand->getSubReg();
+  auto MBB = Mov.getParent();
+  const SIRegisterInfo &TRI = TII->getRegisterInfo();
+
+  // Make sure that 'R' is not modified between the MOVs in 'MBB'.
+  auto I = std::next(Mov.getReverseIterator()), E = MBB->instr_rend();
+  for (; I != E; ++I) {
+    if (instModifiesReg(&*I, R, Rsub, TRI))
+      return false;
+  }
+
+  unsigned Op = Mov.getOpcode();
+  auto SrcOperand = TII->getNamedOperand(Mov, AMDGPU::OpName::src0);
+  assert(SrcOperand->isImm());
+  const int64_t Imm = SrcOperand->getImm();
+
+  // Look for the same MOV in the predecessor.
+  auto Pred = *MBB->pred_begin();
+  I = Pred->instr_rbegin(), E = Pred->instr_rend();
+  for (; I != E; ++I) {
+
+    MachineInstr *Instr = &*I;
+    auto InstrSrcOp = TII->getNamedOperand(*Instr, AMDGPU::OpName::src0);
+    auto InstrDstOp = Instr->getOpcode() == AMDGPU::V_MOV_B32_e32
+                          ? TII->getNamedOperand(*Instr, AMDGPU::OpName::vdst)
+                          : TII->getNamedOperand(*Instr, AMDGPU::OpName::sdst);
+
+    if (Instr->getOpcode() == Op &&
+        InstrDstOp->getReg() == R &&
+        InstrDstOp->getSubReg() == Rsub &&
+        InstrSrcOp->isImm() &&
+        InstrSrcOp->getImm() == Imm) {
+      break;
+    }
+
+    // Make sure that 'R' is not modified between the MOVs in 'Pred'.
+    if (instModifiesReg(Instr, R, Rsub, TRI))
+      return false;
+  }
+
+  // If the same MOV was not found in the predecessor, bail out.
+  if (I == E)
+    return false;
+
+  if (MRI.tracksLiveness() && !MBB->isLiveIn(R))
+    MBB->addLiveIn(R);
+
+  return true;
+}
+
+// Match:
 // mov t, x
 // mov x, y
 // mov y, t
 //
 // =>
 //
 // mov t, x (t is potentially dead and move eliminated)
 // v_swap_b32 x, y
for (I = MBB.begin(); I != MBB.end(); I = Next) {
       if (ST.hasSwap() && (MI.getOpcode() == AMDGPU::V_MOV_B32_e32 ||
                            MI.getOpcode() == AMDGPU::COPY)) {
         if (auto *NextMI = matchSwap(MI, MRI, TII)) {
           Next = NextMI->getIterator();
           continue;
         }
       }
 
+      if (MI.getOpcode() == AMDGPU::V_MOV_B32_e32 ||
+          MI.getOpcode() == AMDGPU::S_MOV_B32 ||
+          MI.getOpcode() == AMDGPU::S_MOV_B64) {
+        // If the MOV is identical to a MOV in the immediate predecessor
+        // of MBB and also no instruction between them modifies the destination
+        // register, then remove the MOV.
+        MachineOperand &Src = MI.getOperand(1);
+        if (Src.isImm() &&
+            TargetRegisterInfo::isPhysicalRegister(MI.getOperand(0).getReg())) {
+
+          if (MBB.pred_size() == 1 && isMovRedundant(MI, MRI, TII)) {
+            MI.eraseFromParent();
+            continue;
+          }
+        }
+      }
+
       // Combine adjacent s_nops to use the immediate operand encoding how long
       // to wait.
       //
       // s_nop N
       // s_nop M
       //  =>
       // s_nop (N + M)
       if (MI.getOpcode() == AMDGPU::S_NOP &&

test/CodeGen/AMDGPU/control-flow-fastregalloc.ll

 ; VMEM: v_mov_b32_e32 v[[V_SAVEEXEC_HI:[0-9]+]], s[[SAVEEXEC_HI]]
 ; VMEM: buffer_store_dword v[[V_SAVEEXEC_HI]], off, s[0:3], s7 offset:24 ; 4-byte Folded Spill
 
 ; GCN: s_mov_b64 exec, s{{\[}}[[ANDEXEC_LO]]:[[ANDEXEC_HI]]{{\]}}
 
 ; GCN: mask branch [[ENDIF:BB[0-9]+_[0-9]+]]
 
 ; GCN: {{^}}BB{{[0-9]+}}_1: ; %if
-; GCN: s_mov_b32 m0, -1
 ; GCN: ds_read_b32 [[LOAD1:v[0-9]+]]
 ; GCN: buffer_load_dword [[RELOAD_LOAD0:v[0-9]+]], off, s[0:3], s7 offset:[[LOAD0_OFFSET]] ; 4-byte Folded Reload
 ; GCN: s_waitcnt vmcnt(0) lgkmcnt(0)
 
 
 ; Spill val register
 ; GCN: v_add_i32_e32 [[VAL:v[0-9]+]], vcc, [[LOAD1]], [[RELOAD_LOAD0]]
 ; GCN: buffer_store_dword [[VAL]], off, s[0:3], s7 offset:[[VAL_OFFSET:[0-9]+]] ; 4-byte Folded Spill

test/CodeGen/AMDGPU/remove-redundant-mov.mir

+# RUN: llc -march=amdgcn -mcpu=gfx900 -run-pass si-shrink-instructions -verify-machineinstrs %s -o - | FileCheck -check-prefix=GCN %s
+
+# GCN-LABEL: name: v_mov_redundant_move_single
+# GCN: bb.1:
+# GCN-NOT: $vgpr2 = V_MOV_B32_e32 1065353216
+# GCN: $vgpr3 = V_MOV_B32_e32 $vgpr2
+---
+name:            v_mov_redundant_move_single
+body:             |
+  bb.0:
+    renamable $vgpr2 = V_MOV_B32_e32 1065353216, implicit $exec
+    S_BRANCH %bb.1
+
+  bb.1:
+    renamable $vgpr2 = V_MOV_B32_e32 1065353216, implicit $exec
+    $vgpr3 = V_MOV_B32_e32 $vgpr2, implicit $exec, implicit $exec
+...
+
+# GCN-LABEL: name: v_mov_redundant_move_liveness
+# GCN: bb.1:
+# GCN: liveins: $vgpr2
+# GCN-NOT: $vgpr2 = V_MOV_B32_e32 1065353216
+# GCN: $vgpr3 = V_MOV_B32_e32 $vgpr2
+---
+name:            v_mov_redundant_move_liveness
+tracksRegLiveness: true
+body:             |
+  bb.0:
+    renamable $vgpr2 = V_MOV_B32_e32 1065353216, implicit $exec
+    S_BRANCH %bb.1
+
+  bb.1:
+    renamable $vgpr2 = V_MOV_B32_e32 1065353216, implicit $exec
+    $vgpr3 = V_MOV_B32_e32 $vgpr2, implicit $exec, implicit $exec
+...
+
+# GCN-LABEL: name: v_mov_redundant_move_multiple
+# GCN: bb.1:
+# GCN-NOT: $vgpr2 = V_MOV_B32_e32 0
+# GCN-NOT: $vgpr1 = V_MOV_B32_e32 1065353216
+# GCN-NOT: $vgpr4 = V_MOV_B32_e32 3204448256
+# GCN-NOT: $vgpr3 = V_MOV_B32_e32 1056964608
+# GCN: $vgpr3 = V_MOV_B32_e32 $vgpr2
+---
+name:            v_mov_redundant_move_multiple
+body:             |
+  bb.0:
+    renamable $vgpr1 = V_MOV_B32_e32 1065353216, implicit $exec
+    renamable $vgpr2 = V_MOV_B32_e32 0, implicit $exec
+    renamable $vgpr3 = V_MOV_B32_e32 1056964608, implicit $exec
+    renamable $vgpr4 = V_MOV_B32_e32 3204448256, implicit $exec
+    S_BRANCH %bb.1
+
+  bb.1:
+    renamable $vgpr2 = V_MOV_B32_e32 0, implicit $exec
+    renamable $vgpr1 = V_MOV_B32_e32 1065353216, implicit $exec
+    renamable $vgpr4 = V_MOV_B32_e32 3204448256, implicit $exec
+    renamable $vgpr3 = V_MOV_B32_e32 1056964608, implicit $exec
+    $vgpr3 = V_MOV_B32_e32 $vgpr2, implicit $exec, implicit $exec
+...
+
+# GCN-LABEL: name: v_mov_necessary_move_not_removed_1
+# GCN: bb.1:
+# GCN: $vgpr2 = V_MOV_B32_e32 1065353216
+# GCN: $vgpr3 = V_MOV_B32_e32 $vgpr2
+---
+name:            v_mov_necessary_move_not_removed_1
+body:             |
+  bb.0:
+    renamable $vgpr2 = V_MOV_B32_e32 1065353216, implicit $exec
+    renamable $vgpr2 = V_MOV_B32_e32 0, implicit $exec
+    S_BRANCH %bb.1
+
+  bb.1:
+    renamable $vgpr2 = V_MOV_B32_e32 1065353216, implicit $exec
+    $vgpr3 = V_MOV_B32_e32 $vgpr2, implicit $exec, implicit $exec
+...
+
+# GCN-LABEL: name: v_mov_necessary_move_not_removed_2
+# GCN: bb.1:
+# GCN: $vgpr2 = V_MOV_B32_e32 1065353216
+# GCN: $vgpr3 = V_MOV_B32_e32 $vgpr2
+---
+name:            v_mov_necessary_move_not_removed_2
+body:             |
+  bb.0:
+    renamable $vgpr2 = V_MOV_B32_e32 1065353216, implicit $exec
+    S_BRANCH %bb.1
+
+  bb.1:
+    renamable $vgpr2 = V_MOV_B32_e32 0, implicit $exec
+    renamable $vgpr2 = V_MOV_B32_e32 1065353216, implicit $exec
+    $vgpr3 = V_MOV_B32_e32 $vgpr2, implicit $exec, implicit $exec
+...
+
+
+
+# GCN-LABEL: name: s_mov_32_redundant_move_single
+# GCN: bb.1:
+# GCN-NOT: $sgpr11 = S_MOV_B32 1065353216
+# GCN: $sgpr12 = S_MOV_B32 $sgpr11
+---
+name:            s_mov_32_redundant_move_single
+body:             |
+  bb.0:
+    renamable $sgpr11 = S_MOV_B32 1065353216, implicit $exec
+    S_BRANCH %bb.1
+
+  bb.1:
+    renamable $sgpr11 = S_MOV_B32 1065353216, implicit $exec
+    $sgpr12 = S_MOV_B32 $sgpr11, implicit $exec, implicit $exec
+...
+
+# GCN-LABEL: name: s_mov_32_redundant_move_liveness
+# GCN: bb.1:
+# GCN: liveins: $sgpr11
+# GCN-NOT: $sgpr11 = S_MOV_B32 1065353216
+# GCN: $sgpr12 = S_MOV_B32 $sgpr11
+---
+name:            s_mov_32_redundant_move_liveness
+tracksRegLiveness: true
+body:             |
+  bb.0:
+    renamable $sgpr11 = S_MOV_B32 1065353216, implicit $exec
+    S_BRANCH %bb.1
+
+  bb.1:
+    renamable $sgpr11 = S_MOV_B32 1065353216, implicit $exec
+    $sgpr12 = S_MOV_B32 $sgpr11, implicit $exec, implicit $exec
+...
+
+# GCN-LABEL: name: s_mov_32_redundant_move_multiple
+# GCN: bb.1:
+# GCN-NOT: $sgpr11 = S_MOV_B32 0
+# GCN-NOT: $sgpr10 = S_MOV_B32 1065353216
+# GCN-NOT: $sgpr13 = S_MOV_B32 3204448256
+# GCN-NOT: $sgpr12 = S_MOV_B32 1056964608
+# GCN: $sgpr12 = S_MOV_B32 $sgpr11
+---
+name:            s_mov_32_redundant_move_multiple
+body:             |
+  bb.0:
+    renamable $sgpr10 = S_MOV_B32 1065353216, implicit $exec
+    renamable $sgpr11 = S_MOV_B32 0, implicit $exec
+    renamable $sgpr12 = S_MOV_B32 1056964608, implicit $exec
+    renamable $sgpr13 = S_MOV_B32 3204448256, implicit $exec
+    S_BRANCH %bb.1
+
+  bb.1:
+    renamable $sgpr11 = S_MOV_B32 0, implicit $exec
+    renamable $sgpr10 = S_MOV_B32 1065353216, implicit $exec
+    renamable $sgpr13 = S_MOV_B32 3204448256, implicit $exec
+    renamable $sgpr12 = S_MOV_B32 1056964608, implicit $exec
+    $sgpr12 = S_MOV_B32 $sgpr11, implicit $exec, implicit $exec
+...
+
+# GCN-LABEL: name: s_mov_32_necessary_move_not_removed_1
+# GCN: bb.1:
+# GCN: $sgpr11 = S_MOV_B32 1065353216
+# GCN: $sgpr12 = S_MOV_B32 $sgpr11
+---
+name:            s_mov_32_necessary_move_not_removed_1
+body:             |
+  bb.0:
+    renamable $sgpr11 = S_MOV_B32 1065353216, implicit $exec
+    renamable $sgpr11 = S_MOV_B32 0, implicit $exec
+    S_BRANCH %bb.1
+
+  bb.1:
+    renamable $sgpr11 = S_MOV_B32 1065353216, implicit $exec
+    $sgpr12 = S_MOV_B32 $sgpr11, implicit $exec, implicit $exec
+...
+
+# GCN-LABEL: name: s_mov_32_necessary_move_not_removed_2
+# GCN: bb.1:
+# GCN: $sgpr11 = S_MOV_B32 1065353216
+# GCN: $sgpr12 = S_MOV_B32 $sgpr11
+---
+name:            s_mov_32_necessary_move_not_removed_2
+body:             |
+  bb.0:
+    renamable $sgpr11 = S_MOV_B32 1065353216, implicit $exec
+    S_BRANCH %bb.1
+
+  bb.1:
+    renamable $sgpr11 = S_MOV_B32 0, implicit $exec
+    renamable $sgpr11 = S_MOV_B32 1065353216, implicit $exec
+    $sgpr12 = S_MOV_B32 $sgpr11, implicit $exec, implicit $exec
+...
+
+
+
+# GCN-LABEL: name: s_mov_64_redundant_move_single
+# GCN: bb.1:
+# GCN-NOT: $sgpr8_sgpr9 = S_MOV_B64 1065353216
+# GCN: $sgpr6_sgpr7 = S_MOV_B64 $sgpr8_sgpr9
+---
+name:            s_mov_64_redundant_move_single
+body:             |
+  bb.0:
+    renamable $sgpr8_sgpr9 = S_MOV_B64 1065353216, implicit $exec
+    S_BRANCH %bb.1
+
+  bb.1:
+    renamable $sgpr8_sgpr9 = S_MOV_B64 1065353216, implicit $exec
+    $sgpr6_sgpr7 = S_MOV_B64 $sgpr8_sgpr9, implicit $exec, implicit $exec
+...
+
+# GCN-LABEL: name: s_mov_64_redundant_move_liveness
+# GCN: bb.1:
+# GCN: liveins: $sgpr10_sgpr11
+# GCN-NOT: $sgpr10_sgpr11 = S_MOV_B64 1065353216
+# GCN: $sgpr6_sgpr7 = S_MOV_B64 $sgpr10_sgpr11
+---
+name:            s_mov_64_redundant_move_liveness
+tracksRegLiveness: true
+body:             |
+  bb.0:
+    renamable $sgpr10_sgpr11 = S_MOV_B64 1065353216, implicit $exec
+    S_BRANCH %bb.1
+
+  bb.1:
+    renamable $sgpr10_sgpr11 = S_MOV_B64 1065353216, implicit $exec
+    $sgpr6_sgpr7 = S_MOV_B64 $sgpr10_sgpr11, implicit $exec, implicit $exec
+...
+
+# GCN-LABEL: name: s_mov_64_redundant_move_multiple
+# GCN: bb.1:
+# GCN-NOT: $sgpr10_sgpr11 = S_MOV_B64 0
+# GCN-NOT: $sgpr12_sgpr13 = S_MOV_B64 1065353216
+# GCN-NOT: $sgpr6_sgpr7 = S_MOV_B64 3204448256
+# GCN-NOT: $sgpr8_sgpr9 = S_MOV_B64 1056964608
+# GCN: $sgpr14_sgpr15 = S_MOV_B64 $sgpr8_sgpr9
+---
+name:            s_mov_64_redundant_move_multiple
+body:             |
+  bb.0:
+    renamable $sgpr12_sgpr13 = S_MOV_B64 1065353216, implicit $exec
+    renamable $sgpr10_sgpr11 = S_MOV_B64 0, implicit $exec
+    renamable $sgpr8_sgpr9 = S_MOV_B64 1056964608, implicit $exec
+    renamable $sgpr6_sgpr7 = S_MOV_B64 3204448256, implicit $exec
+    S_BRANCH %bb.1
+
+  bb.1:
+    renamable $sgpr10_sgpr11 = S_MOV_B64 0, implicit $exec
+    renamable $sgpr12_sgpr13 = S_MOV_B64 1065353216, implicit $exec
+    renamable $sgpr6_sgpr7 = S_MOV_B64 3204448256, implicit $exec
+    renamable $sgpr8_sgpr9 = S_MOV_B64 1056964608, implicit $exec
+    $sgpr14_sgpr15 = S_MOV_B64 $sgpr8_sgpr9, implicit $exec, implicit $exec
+...
+
+# GCN-LABEL: name: s_mov_64_necessary_move_not_removed_1
+# GCN: bb.1:
+# GCN: $sgpr10_sgpr11 = S_MOV_B64 1065353216
+# GCN: $sgpr6_sgpr7 = S_MOV_B64 $sgpr10_sgpr11
+---
+name:            s_mov_64_necessary_move_not_removed_1
+body:             |
+  bb.0:
+    renamable $sgpr10_sgpr11 = S_MOV_B64 1065353216, implicit $exec
+    renamable $sgpr10_sgpr11 = S_MOV_B64 0, implicit $exec
+    S_BRANCH %bb.1
+
+  bb.1:
+    renamable $sgpr10_sgpr11 = S_MOV_B64 1065353216, implicit $exec
+    $sgpr6_sgpr7 = S_MOV_B64 $sgpr10_sgpr11, implicit $exec, implicit $exec
+...
+
+# GCN-LABEL: name: s_mov_64_necessary_move_not_removed_2
+# GCN: bb.1:
+# GCN: $sgpr10_sgpr11 = S_MOV_B64 1065353216
+# GCN: $sgpr6_sgpr7 = S_MOV_B64 $sgpr10_sgpr11
+---
+name:            s_mov_64_necessary_move_not_removed_2
+body:             |
+  bb.0:
+    renamable $sgpr10_sgpr11 = S_MOV_B64 1065353216, implicit $exec
+    S_BRANCH %bb.1
+
+  bb.1:
+    renamable $sgpr10_sgpr11 = S_MOV_B64 0, implicit $exec
+    renamable $sgpr10_sgpr11 = S_MOV_B64 1065353216, implicit $exec
+    $sgpr6_sgpr7 = S_MOV_B64 $sgpr10_sgpr11, implicit $exec, implicit $exec
+...