[AIE2] Support for 64-bit G_BUILD_VECTOR and G_UNMERGE_VALUES

Xilinx · Jun 25, 2024 · 226d716 · 226d716
1 parent 118f973
commit 226d716
Show file tree

Hide file tree

Showing 4 changed files with 253 additions and 34 deletions.
diff --git a/llvm/lib/Target/AIE/AIELegalizerInfo.cpp b/llvm/lib/Target/AIE/AIELegalizerInfo.cpp
@@ -89,7 +89,7 @@ static LegalityPredicate isValidVectorAIE2(const unsigned TypeIdx) {
   return [=](const LegalityQuery &Query) {
     const LLT DstTy = Query.Types[TypeIdx];
     const unsigned DstSize = DstTy.getSizeInBits();
-    return DstTy.isVector() && (DstSize == 32 || DstSize > 64);
+    return DstTy.isVector() && (DstSize >= 32);
   };
 }
 
@@ -551,44 +551,46 @@ bool AIELegalizerInfo::legalizeCustom(LegalizerHelper &Helper,
   llvm_unreachable("Un-expected custom legalization");
 }
 
-bool AIELegalizerInfo::pack32BitVector(LegalizerHelper &Helper,
-                                       MachineInstr &MI,
-                                       Register SourceReg) const {
+bool AIELegalizerInfo::packBitVector(LegalizerHelper &Helper,
+                                     MachineInstr &MI) const {
   MachineIRBuilder &MIRBuilder = Helper.MIRBuilder;
   MachineRegisterInfo &MRI = *MIRBuilder.getMRI();
 
-  const LLT SourceRegTy = MRI.getType(SourceReg);
   const Register DstReg = MI.getOperand(0).getReg();
-  assert(SourceRegTy.getSizeInBits() == 32 &&
-         "cannot pack vectors larger or smaller than 32-bit");
+  const LLT DstRegTy = MRI.getType(DstReg);
+  assert(DstRegTy.getSizeInBits() == 32 ||
+         DstRegTy.getSizeInBits() == 64 &&
+             "can only pack 32- and 64-bit vectors");
 
-  const LLT S32 = LLT::scalar(32);
+  const LLT STy = LLT::scalar(MRI.getType(DstReg).getSizeInBits());
   unsigned Offset = 0;
-  Register DstCastReg = MRI.createGenericVirtualRegister(S32);
+  Register DstCastReg = MRI.createGenericVirtualRegister(STy);
 
   // Skip the destination operand since that is where we are writing to.
   MachineOperand *Operand = MI.operands_begin() + 1,
                  *OperandEnd = MI.operands_end();
   MIRBuilder.buildConstant(DstCastReg, 0);
 
-  const LLT RegTy = MRI.getType(DstReg);
+  const LLT RegTy = MRI.getType(Operand->getReg());
   while (Operand != OperandEnd) {
     Register DestinationOperand = Operand->getReg();
 
-    if (RegTy.getScalarSizeInBits() != 32) {
-      const Register TmpReg32 = MRI.createGenericVirtualRegister(S32);
-      MIRBuilder.buildInstr(AIE2::G_ZEXT, {TmpReg32}, {DestinationOperand});
-      DestinationOperand = TmpReg32;
+    const LLT S32 = LLT::scalar(32);
+    if (RegTy != STy) {
+      const Register TmpReg = MRI.createGenericVirtualRegister(STy);
+      MIRBuilder.buildInstr(AIE2::G_ZEXT, {TmpReg}, {DestinationOperand});
+      DestinationOperand = TmpReg;
     }
 
     // Avoid a useless shift for the first element, since it doesn't get
     // optimized out in O0.
-    const Register AccumulatorReg = MRI.createGenericVirtualRegister(S32);
+    const Register AccumulatorReg = MRI.createGenericVirtualRegister(STy);
+
     if (Offset != 0) {
       const MachineInstrBuilder ShiftConstant =
           MIRBuilder.buildConstant(S32, Offset);
       const MachineInstrBuilder Masked =
-          MIRBuilder.buildShl(S32, DestinationOperand, ShiftConstant);
+          MIRBuilder.buildShl(STy, DestinationOperand, ShiftConstant);
       MIRBuilder.buildOr(AccumulatorReg, DstCastReg, Masked);
     } else {
       MIRBuilder.buildOr(AccumulatorReg, DstCastReg, DestinationOperand);
@@ -604,19 +606,20 @@ bool AIELegalizerInfo::pack32BitVector(LegalizerHelper &Helper,
   return true;
 }
 
-bool AIELegalizerInfo::unpack32BitVector(LegalizerHelper &Helper,
-                                         MachineInstr &MI,
-                                         Register SourceReg) const {
+bool AIELegalizerInfo::unpackBitVector(LegalizerHelper &Helper,
+                                       MachineInstr &MI) const {
   MachineIRBuilder &MIRBuilder = Helper.MIRBuilder;
   MachineRegisterInfo &MRI = *MIRBuilder.getMRI();
 
+  const Register SourceReg = MI.getOperand(MI.getNumOperands() - 1).getReg();
   const LLT SourceRegTy = MRI.getType(SourceReg);
-  assert(SourceRegTy.getSizeInBits() == 32 &&
-         "cannot unpack vectors larger or smaller than 32-bit");
+  assert(SourceRegTy.getSizeInBits() == 32 ||
+         SourceRegTy.getSizeInBits() == 64 &&
+             "can only unpack 32- and 64-bit vectors");
 
-  const LLT S32 = LLT::scalar(32);
   unsigned Offset = 0;
-  Register DstCastReg = MRI.createGenericVirtualRegister(S32);
+  const LLT STy = LLT::scalar(SourceRegTy.getSizeInBits());
+  Register DstCastReg = MRI.createGenericVirtualRegister(STy);
 
   MachineOperand *Operand = MI.operands_begin(),
                  *OperandEnd = MI.operands_end() - 1;
@@ -627,10 +630,11 @@ bool AIELegalizerInfo::unpack32BitVector(LegalizerHelper &Helper,
     // Avoid a useless shift for the first element, since it doesn't get
     // optimized out in O0.
     if (Offset != 0) {
+      const LLT S32 = LLT::scalar(32);
       const MachineInstrBuilder ShiftConstant =
           MIRBuilder.buildConstant(S32, Offset);
       const MachineInstrBuilder Masked =
-          MIRBuilder.buildLShr(S32, DstCastReg, ShiftConstant);
+          MIRBuilder.buildLShr(STy, DstCastReg, ShiftConstant);
       MIRBuilder.buildTrunc(DestinationOperand, Masked);
 
     } else {
@@ -658,13 +662,13 @@ bool AIELegalizerInfo::legalizeG_BUILD_VECTOR(LegalizerHelper &Helper,
   const unsigned EltSize = DstVecEltTy.getScalarSizeInBits();
   assert((EltSize == 8 || EltSize == 16 || EltSize == 32) &&
          "non-existent integer size");
-  assert(DstVecSize == 32 || (DstVecSize > 64 && DstVecSize <= 1024 &&
-                              "non-native vectors are not supported"));
+  assert(DstVecSize >= 32 && DstVecSize <= 1024 &&
+         "non-native vectors are not supported");
   assert(DstVecSize < 1024 && "vadd takes a 512-bit argument");
 
   // If our vector is 32-bit we can store it as packed integer vector
-  if (DstVecSize == 32)
-    return pack32BitVector(Helper, MI, DstReg);
+  if (DstVecSize == 32 || DstVecSize == 64)
+    return packBitVector(Helper, MI);
 
   // We are using an undef since we are building over multiple instructions
   const TypeSize VecEltTySize = DstVecEltTy.getSizeInBits();
@@ -724,8 +728,8 @@ bool AIELegalizerInfo::legalizeG_UNMERGE_VALUES(LegalizerHelper &Helper,
              LastTy.getSizeInBits() &&
          "This operation is only supported for vectors");
 
-  if (LastTy.getSizeInBits() == 32)
-    return unpack32BitVector(Helper, MI, LastReg);
+  if (LastTy.getSizeInBits() == 32 || LastTy.getSizeInBits() == 64)
+    return unpackBitVector(Helper, MI);
 
   // Pad vectors of 128-bit vectors to 256-bit
   Register TargetReg = LastReg;

diff --git a/llvm/lib/Target/AIE/AIELegalizerInfo.h b/llvm/lib/Target/AIE/AIELegalizerInfo.h
@@ -51,10 +51,8 @@ class AIELegalizerInfo : public LegalizerInfo {
   bool legalizeG_FADDSUB(LegalizerHelper &Helper, MachineInstr &MI) const;
 
   // Helper functions for legalization
-  bool pack32BitVector(LegalizerHelper &Helper, MachineInstr &MI,
-                       Register SourceReg) const;
-  bool unpack32BitVector(LegalizerHelper &Helper, MachineInstr &MI,
-                         Register SourceReg) const;
+  bool packBitVector(LegalizerHelper &Helper, MachineInstr &MI) const;
+  bool unpackBitVector(LegalizerHelper &Helper, MachineInstr &MI) const;
 };
 } // end namespace llvm
 #endif
diff --git a/llvm/test/CodeGen/AIE/aie2/GlobalISel/legalize-build-vector.mir b/llvm/test/CodeGen/AIE/aie2/GlobalISel/legalize-build-vector.mir
@@ -661,3 +661,111 @@ body:             |
     %5:_(s8) = G_TRUNC %1(s32)
     %6:_(<4 x s8>) = G_BUILD_VECTOR %2(s8), %3(s8), %4(s8), %5(s8)
     PseudoRET implicit $lr, implicit %6
+...
+
+---
+name:            test_build_vector_64_16
+body:             |
+  bb.1.entry:
+    ; CHECK-LABEL: name: test_build_vector_64_16
+    ; CHECK: [[C:%[0-9]+]]:_(s32) = G_CONSTANT i32 1
+    ; CHECK-NEXT: [[C1:%[0-9]+]]:_(s32) = G_CONSTANT i32 2
+    ; CHECK-NEXT: [[C2:%[0-9]+]]:_(s32) = G_CONSTANT i32 3
+    ; CHECK-NEXT: [[C3:%[0-9]+]]:_(s32) = G_CONSTANT i32 4
+    ; CHECK-NEXT: [[C4:%[0-9]+]]:_(s32) = G_CONSTANT i32 0
+    ; CHECK-NEXT: [[DEF:%[0-9]+]]:_(s32) = G_IMPLICIT_DEF
+    ; CHECK-NEXT: [[C5:%[0-9]+]]:_(s32) = G_CONSTANT i32 65535
+    ; CHECK-NEXT: [[AND:%[0-9]+]]:_(s32) = G_AND [[C]], [[C5]]
+    ; CHECK-NEXT: [[AND1:%[0-9]+]]:_(s32) = G_AND [[DEF]], [[C4]]
+    ; CHECK-NEXT: [[OR:%[0-9]+]]:_(s32) = G_OR [[C4]], [[AND]]
+    ; CHECK-NEXT: [[OR1:%[0-9]+]]:_(s32) = G_OR [[C4]], [[AND1]]
+    ; CHECK-NEXT: [[AND2:%[0-9]+]]:_(s32) = G_AND [[C1]], [[C5]]
+    ; CHECK-NEXT: [[AND3:%[0-9]+]]:_(s32) = G_AND [[DEF]], [[C4]]
+    ; CHECK-NEXT: [[C6:%[0-9]+]]:_(s32) = G_CONSTANT i32 16
+    ; CHECK-NEXT: [[SHL:%[0-9]+]]:_(s32) = G_SHL [[AND2]], [[C6]](s32)
+    ; CHECK-NEXT: [[SHL1:%[0-9]+]]:_(s32) = G_SHL [[AND3]], [[C6]](s32)
+    ; CHECK-NEXT: [[LSHR:%[0-9]+]]:_(s32) = G_LSHR [[AND2]], [[C6]](s32)
+    ; CHECK-NEXT: [[OR2:%[0-9]+]]:_(s32) = G_OR [[SHL1]], [[LSHR]]
+    ; CHECK-NEXT: [[OR3:%[0-9]+]]:_(s32) = G_OR [[OR]], [[SHL]]
+    ; CHECK-NEXT: [[OR4:%[0-9]+]]:_(s32) = G_OR [[OR1]], [[OR2]]
+    ; CHECK-NEXT: [[AND4:%[0-9]+]]:_(s32) = G_AND [[C2]], [[C5]]
+    ; CHECK-NEXT: [[OR5:%[0-9]+]]:_(s32) = G_OR [[OR3]], [[C4]]
+    ; CHECK-NEXT: [[OR6:%[0-9]+]]:_(s32) = G_OR [[OR4]], [[AND4]]
+    ; CHECK-NEXT: [[AND5:%[0-9]+]]:_(s32) = G_AND [[C3]], [[C5]]
+    ; CHECK-NEXT: [[SHL2:%[0-9]+]]:_(s32) = G_SHL [[AND5]], [[C6]](s32)
+    ; CHECK-NEXT: [[OR7:%[0-9]+]]:_(s32) = G_OR [[OR5]], [[C4]]
+    ; CHECK-NEXT: [[OR8:%[0-9]+]]:_(s32) = G_OR [[OR6]], [[SHL2]]
+    ; CHECK-NEXT: [[MV:%[0-9]+]]:_(s64) = G_MERGE_VALUES [[OR7]](s32), [[OR8]](s32)
+    ; CHECK-NEXT: [[BITCAST:%[0-9]+]]:_(<4 x s16>) = G_BITCAST [[MV]](s64)
+    ; CHECK-NEXT: PseudoRET implicit $lr, implicit [[BITCAST]](<4 x s16>)
+    %0:_(s16) = G_CONSTANT i16 1
+    %1:_(s16) = G_CONSTANT i16 2
+    %2:_(s16) = G_CONSTANT i16 3
+    %3:_(s16) = G_CONSTANT i16 4
+    %4:_(<4 x s16>) = G_BUILD_VECTOR %0(s16), %1(s16), %2(s16), %3(s16)
+    PseudoRET implicit $lr, implicit %4
+...
+
+---
+name:            test_build_vector_64_8
+body:             |
+  bb.1.entry:
+    ; CHECK-LABEL: name: test_build_vector_64_8
+    ; CHECK: [[C:%[0-9]+]]:_(s32) = G_CONSTANT i32 1
+    ; CHECK-NEXT: [[C1:%[0-9]+]]:_(s32) = G_CONSTANT i32 2
+    ; CHECK-NEXT: [[C2:%[0-9]+]]:_(s32) = G_CONSTANT i32 3
+    ; CHECK-NEXT: [[C3:%[0-9]+]]:_(s32) = G_CONSTANT i32 4
+    ; CHECK-NEXT: [[C4:%[0-9]+]]:_(s32) = G_CONSTANT i32 0
+    ; CHECK-NEXT: [[DEF:%[0-9]+]]:_(s32) = G_IMPLICIT_DEF
+    ; CHECK-NEXT: [[C5:%[0-9]+]]:_(s32) = G_CONSTANT i32 255
+    ; CHECK-NEXT: [[AND:%[0-9]+]]:_(s32) = G_AND [[C]], [[C5]]
+    ; CHECK-NEXT: [[AND1:%[0-9]+]]:_(s32) = G_AND [[DEF]], [[C4]]
+    ; CHECK-NEXT: [[OR:%[0-9]+]]:_(s32) = G_OR [[C4]], [[AND]]
+    ; CHECK-NEXT: [[OR1:%[0-9]+]]:_(s32) = G_OR [[C4]], [[AND1]]
+    ; CHECK-NEXT: [[AND2:%[0-9]+]]:_(s32) = G_AND [[C1]], [[C5]]
+    ; CHECK-NEXT: [[AND3:%[0-9]+]]:_(s32) = G_AND [[DEF]], [[C4]]
+    ; CHECK-NEXT: [[C6:%[0-9]+]]:_(s32) = G_CONSTANT i32 8
+    ; CHECK-NEXT: [[SHL:%[0-9]+]]:_(s32) = G_SHL [[AND2]], [[C6]](s32)
+    ; CHECK-NEXT: [[SHL1:%[0-9]+]]:_(s32) = G_SHL [[AND3]], [[C6]](s32)
+    ; CHECK-NEXT: [[C7:%[0-9]+]]:_(s32) = G_CONSTANT i32 24
+    ; CHECK-NEXT: [[LSHR:%[0-9]+]]:_(s32) = G_LSHR [[AND2]], [[C7]](s32)
+    ; CHECK-NEXT: [[OR2:%[0-9]+]]:_(s32) = G_OR [[SHL1]], [[LSHR]]
+    ; CHECK-NEXT: [[OR3:%[0-9]+]]:_(s32) = G_OR [[OR]], [[SHL]]
+    ; CHECK-NEXT: [[OR4:%[0-9]+]]:_(s32) = G_OR [[OR1]], [[OR2]]
+    ; CHECK-NEXT: [[AND4:%[0-9]+]]:_(s32) = G_AND [[C2]], [[C5]]
+    ; CHECK-NEXT: [[AND5:%[0-9]+]]:_(s32) = G_AND [[DEF]], [[C4]]
+    ; CHECK-NEXT: [[C8:%[0-9]+]]:_(s32) = G_CONSTANT i32 16
+    ; CHECK-NEXT: [[SHL2:%[0-9]+]]:_(s32) = G_SHL [[AND4]], [[C8]](s32)
+    ; CHECK-NEXT: [[SHL3:%[0-9]+]]:_(s32) = G_SHL [[AND5]], [[C8]](s32)
+    ; CHECK-NEXT: [[LSHR1:%[0-9]+]]:_(s32) = G_LSHR [[AND4]], [[C8]](s32)
+    ; CHECK-NEXT: [[OR5:%[0-9]+]]:_(s32) = G_OR [[SHL3]], [[LSHR1]]
+    ; CHECK-NEXT: [[OR6:%[0-9]+]]:_(s32) = G_OR [[OR3]], [[SHL2]]
+    ; CHECK-NEXT: [[OR7:%[0-9]+]]:_(s32) = G_OR [[OR4]], [[OR5]]
+    ; CHECK-NEXT: [[AND6:%[0-9]+]]:_(s32) = G_AND [[C3]], [[C5]]
+    ; CHECK-NEXT: [[AND7:%[0-9]+]]:_(s32) = G_AND [[DEF]], [[C4]]
+    ; CHECK-NEXT: [[SHL4:%[0-9]+]]:_(s32) = G_SHL [[AND6]], [[C7]](s32)
+    ; CHECK-NEXT: [[SHL5:%[0-9]+]]:_(s32) = G_SHL [[AND7]], [[C7]](s32)
+    ; CHECK-NEXT: [[LSHR2:%[0-9]+]]:_(s32) = G_LSHR [[AND6]], [[C6]](s32)
+    ; CHECK-NEXT: [[OR8:%[0-9]+]]:_(s32) = G_OR [[SHL5]], [[LSHR2]]
+    ; CHECK-NEXT: [[OR9:%[0-9]+]]:_(s32) = G_OR [[OR6]], [[SHL4]]
+    ; CHECK-NEXT: [[OR10:%[0-9]+]]:_(s32) = G_OR [[OR7]], [[OR8]]
+    ; CHECK-NEXT: [[OR11:%[0-9]+]]:_(s32) = G_OR [[OR9]], [[C4]]
+    ; CHECK-NEXT: [[OR12:%[0-9]+]]:_(s32) = G_OR [[OR10]], [[AND6]]
+    ; CHECK-NEXT: [[SHL6:%[0-9]+]]:_(s32) = G_SHL [[AND4]], [[C6]](s32)
+    ; CHECK-NEXT: [[OR13:%[0-9]+]]:_(s32) = G_OR [[OR11]], [[C4]]
+    ; CHECK-NEXT: [[OR14:%[0-9]+]]:_(s32) = G_OR [[OR12]], [[SHL6]]
+    ; CHECK-NEXT: [[SHL7:%[0-9]+]]:_(s32) = G_SHL [[AND2]], [[C8]](s32)
+    ; CHECK-NEXT: [[OR15:%[0-9]+]]:_(s32) = G_OR [[OR13]], [[C4]]
+    ; CHECK-NEXT: [[OR16:%[0-9]+]]:_(s32) = G_OR [[OR14]], [[SHL7]]
+    ; CHECK-NEXT: [[SHL8:%[0-9]+]]:_(s32) = G_SHL [[AND]], [[C7]](s32)
+    ; CHECK-NEXT: [[OR17:%[0-9]+]]:_(s32) = G_OR [[OR15]], [[C4]]
+    ; CHECK-NEXT: [[OR18:%[0-9]+]]:_(s32) = G_OR [[OR16]], [[SHL8]]
+    ; CHECK-NEXT: [[MV:%[0-9]+]]:_(s64) = G_MERGE_VALUES [[OR17]](s32), [[OR18]](s32)
+    ; CHECK-NEXT: [[BITCAST:%[0-9]+]]:_(<8 x s8>) = G_BITCAST [[MV]](s64)
+    ; CHECK-NEXT: PseudoRET implicit $lr, implicit [[BITCAST]](<8 x s8>)
+    %0:_(s8) = G_CONSTANT i8 1
+    %1:_(s8) = G_CONSTANT i8 2
+    %2:_(s8) = G_CONSTANT i8 3
+    %3:_(s8) = G_CONSTANT i8 4
+    %4:_(<8 x s8>) = G_BUILD_VECTOR %0(s8), %1(s8), %2(s8), %3(s8), %3(s8), %2(s8), %1(s8), %0(s8)
+    PseudoRET implicit $lr, implicit %4
diff --git a/llvm/test/CodeGen/AIE/aie2/GlobalISel/legalize-unmerge-values.mir b/llvm/test/CodeGen/AIE/aie2/GlobalISel/legalize-unmerge-values.mir
@@ -525,3 +525,112 @@ body:             |
     %1(s32), %2(s32), %3(s32), %4(s32) = G_UNMERGE_VALUES %5(<4 x s32>)
     PseudoRET implicit $lr, implicit %1, implicit %2, implicit %3, implicit %4
 ...
+---
+name:           test_unmerge_v4s16
+registers:
+  - { id: 0, class: _, preferred-register: '' }
+  - { id: 1, class: _, preferred-register: '' }
+  - { id: 2, class: _, preferred-register: '' }
+  - { id: 3, class: _, preferred-register: '' }
+  - { id: 4, class: _, preferred-register: '' }
+  - { id: 5, class: _, preferred-register: '' }
+  - { id: 6, class: _, preferred-register: '' }
+  - { id: 7, class: _, preferred-register: '' }
+  - { id: 8, class: _, preferred-register: '' }
+legalized: false
+body:             |
+  bb.0.entry:
+    liveins: $l0
+    ; CHECK-LABEL: name: test_unmerge_v4s16
+    ; CHECK: liveins: $l0
+    ; CHECK-NEXT: {{  $}}
+    ; CHECK-NEXT: [[COPY:%[0-9]+]]:_(<4 x s16>) = COPY $l0
+    ; CHECK-NEXT: [[BITCAST:%[0-9]+]]:_(s64) = G_BITCAST [[COPY]](<4 x s16>)
+    ; CHECK-NEXT: [[C:%[0-9]+]]:_(s32) = G_CONSTANT i32 16
+    ; CHECK-NEXT: [[UV:%[0-9]+]]:_(s32), [[UV1:%[0-9]+]]:_(s32) = G_UNMERGE_VALUES [[BITCAST]](s64)
+    ; CHECK-NEXT: [[LSHR:%[0-9]+]]:_(s32) = G_LSHR [[UV1]], [[C]](s32)
+    ; CHECK-NEXT: [[SEXT_INREG:%[0-9]+]]:_(s32) = G_SEXT_INREG [[LSHR]], 16
+    ; CHECK-NEXT: PseudoRET implicit $lr, implicit [[SEXT_INREG]](s32)
+    %0:_(<4 x s16>) = COPY $l0
+    %1(s16), %2(s16), %3(s16), %4(s16) = G_UNMERGE_VALUES %0(<4 x s16>)
+    %5(s32) = G_SEXT %4(s16)
+    PseudoRET implicit $lr, implicit %5
+...
+---
+name:           test_unmerge_v8s8
+registers:
+  - { id: 0, class: _, preferred-register: '' }
+  - { id: 1, class: _, preferred-register: '' }
+  - { id: 2, class: _, preferred-register: '' }
+  - { id: 3, class: _, preferred-register: '' }
+  - { id: 4, class: _, preferred-register: '' }
+  - { id: 5, class: _, preferred-register: '' }
+  - { id: 6, class: _, preferred-register: '' }
+  - { id: 7, class: _, preferred-register: '' }
+  - { id: 8, class: _, preferred-register: '' }
+  - { id: 9, class: _, preferred-register: '' }
+
+legalized: false
+body:             |
+  bb.0.entry:
+    liveins: $l0
+    ; CHECK-LABEL: name: test_unmerge_v8s8
+    ; CHECK: liveins: $l0
+    ; CHECK-NEXT: {{  $}}
+    ; CHECK-NEXT: [[COPY:%[0-9]+]]:_(<8 x s8>) = COPY $l0
+    ; CHECK-NEXT: [[BITCAST:%[0-9]+]]:_(s64) = G_BITCAST [[COPY]](<8 x s8>)
+    ; CHECK-NEXT: [[C:%[0-9]+]]:_(s32) = G_CONSTANT i32 8
+    ; CHECK-NEXT: [[C1:%[0-9]+]]:_(s32) = G_CONSTANT i32 24
+    ; CHECK-NEXT: [[UV:%[0-9]+]]:_(s32), [[UV1:%[0-9]+]]:_(s32) = G_UNMERGE_VALUES [[BITCAST]](s64)
+    ; CHECK-NEXT: [[LSHR:%[0-9]+]]:_(s32) = G_LSHR [[UV]], [[C1]](s32)
+    ; CHECK-NEXT: [[SHL:%[0-9]+]]:_(s32) = G_SHL [[UV1]], [[C]](s32)
+    ; CHECK-NEXT: [[OR:%[0-9]+]]:_(s32) = G_OR [[LSHR]], [[SHL]]
+    ; CHECK-NEXT: [[SEXT_INREG:%[0-9]+]]:_(s32) = G_SEXT_INREG [[OR]], 8
+    ; CHECK-NEXT: PseudoRET implicit $lr, implicit [[SEXT_INREG]](s32)
+    %0:_(<8 x s8>) = COPY $l0
+    %1(s8), %2(s8), %3(s8), %4(s8), %5(s8), %6(s8), %7(s8), %8(s8) = G_UNMERGE_VALUES %0(<8 x s8>)
+    %9(s32) = G_SEXT %4(s8)
+    PseudoRET implicit $lr, implicit %9
+...
+---
+name:           test_unmerge_64bit_order
+registers:
+  - { id: 0, class: _, preferred-register: '' }
+  - { id: 1, class: _, preferred-register: '' }
+  - { id: 2, class: _, preferred-register: '' }
+  - { id: 3, class: _, preferred-register: '' }
+  - { id: 4, class: _, preferred-register: '' }
+  - { id: 5, class: _, preferred-register: '' }
+  - { id: 6, class: _, preferred-register: '' }
+  - { id: 7, class: _, preferred-register: '' }
+  - { id: 8, class: _, preferred-register: '' }
+legalized: false
+body:             |
+  bb.0.entry:
+    liveins: $l0
+    ; CHECK-LABEL: name: test_unmerge_64bit_order
+    ; CHECK: liveins: $l0
+    ; CHECK-NEXT: {{  $}}
+    ; CHECK-NEXT: [[COPY:%[0-9]+]]:_(<4 x s16>) = COPY $l0
+    ; CHECK-NEXT: [[BITCAST:%[0-9]+]]:_(s64) = G_BITCAST [[COPY]](<4 x s16>)
+    ; CHECK-NEXT: [[C:%[0-9]+]]:_(s32) = G_CONSTANT i32 16
+    ; CHECK-NEXT: [[UV:%[0-9]+]]:_(s32), [[UV1:%[0-9]+]]:_(s32) = G_UNMERGE_VALUES [[BITCAST]](s64)
+    ; CHECK-NEXT: [[LSHR:%[0-9]+]]:_(s32) = G_LSHR [[UV]], [[C]](s32)
+    ; CHECK-NEXT: [[SHL:%[0-9]+]]:_(s32) = G_SHL [[UV1]], [[C]](s32)
+    ; CHECK-NEXT: [[OR:%[0-9]+]]:_(s32) = G_OR [[LSHR]], [[SHL]]
+    ; CHECK-NEXT: [[UV2:%[0-9]+]]:_(s32), [[UV3:%[0-9]+]]:_(s32) = G_UNMERGE_VALUES [[BITCAST]](s64)
+    ; CHECK-NEXT: [[UV4:%[0-9]+]]:_(s32), [[UV5:%[0-9]+]]:_(s32) = G_UNMERGE_VALUES [[BITCAST]](s64)
+    ; CHECK-NEXT: [[LSHR1:%[0-9]+]]:_(s32) = G_LSHR [[UV5]], [[C]](s32)
+    ; CHECK-NEXT: [[TRUNC:%[0-9]+]]:_(s32) = G_TRUNC [[BITCAST]](s64)
+    ; CHECK-NEXT: [[SEXT_INREG:%[0-9]+]]:_(s32) = G_SEXT_INREG [[TRUNC]], 16
+    ; CHECK-NEXT: [[SEXT_INREG1:%[0-9]+]]:_(s32) = G_SEXT_INREG [[OR]], 16
+    ; CHECK-NEXT: [[SEXT_INREG2:%[0-9]+]]:_(s32) = G_SEXT_INREG [[UV3]], 16
+    ; CHECK-NEXT: [[SEXT_INREG3:%[0-9]+]]:_(s32) = G_SEXT_INREG [[LSHR1]], 16
+    ; CHECK-NEXT: PseudoRET implicit $lr, implicit [[SEXT_INREG]](s32), implicit [[SEXT_INREG1]](s32), implicit [[SEXT_INREG2]](s32), implicit [[SEXT_INREG3]](s32)
+    %0:_(<4 x s16>) = COPY $l0
+    %1(s16), %2(s16), %3(s16), %4(s16) = G_UNMERGE_VALUES %0(<4 x s16>)
+    %5(s32) = G_SEXT %1(s16)
+    %6(s32) = G_SEXT %2(s16)
+    %7(s32) = G_SEXT %3(s16)
+    %8(s32) = G_SEXT %4(s16)
+    PseudoRET implicit $lr, implicit %5, implicit %6, implicit %7, implicit %8