[LoongArch] Improve codegen for i8/i16 'atomicrmw xchg a, {0,-1}'

[hev]

Summary

Similar to D156801 for RISCV.

Link: https://github.com/rust-lang/rust/pull/114034
Link: https://github.com/llvm/llvm-project/issues/64090

[xen0n]

I've checked the provided context and this patch seems a correct port of the desired behavior. Thanks!

llvm/lib/Target/LoongArch/LoongArchISelLowering.cpp

Value *Result = Builder.CreateCall(
       MaskedCmpXchg, {AlignedAddr, CmpVal, NewVal, Mask, Ordering});
   Result = Builder.CreateTrunc(Result, Builder.getInt32Ty());
   return Result;
 }
 
 Value *LoongArchTargetLowering::emitMaskedAtomicRMWIntrinsic(
     IRBuilderBase &Builder, AtomicRMWInst *AI, Value *AlignedAddr, Value *Incr,
     Value *Mask, Value *ShiftAmt, AtomicOrdering Ord) const {
+  // In the case of an atomicrmw xchg with a constant 0/-1 operand, replace
+  // the atomic instruction with an AtomicRMWInst::And/Or with appropriate
+  // mask, as this produces better code than the LL/SC loop emitted by
+  // int_loongarch_masked_atomicrmw_xchg.
+  if (AI->getOperation() == AtomicRMWInst::Xchg &&
+      isa<ConstantInt>(AI->getValOperand())) {
+    ConstantInt *CVal = cast<ConstantInt>(AI->getValOperand());
+    if (CVal->isZero())
+      return Builder.CreateAtomicRMW(AtomicRMWInst::And, AlignedAddr,
+                                     Builder.CreateNot(Mask, "Inv_Mask"),
+                                     AI->getAlign(), Ord);
+    if (CVal->isMinusOne())
+      return Builder.CreateAtomicRMW(AtomicRMWInst::Or, AlignedAddr, Mask,
+                                     AI->getAlign(), Ord);
+  }
+
   unsigned GRLen = Subtarget.getGRLen();
   Value *Ordering =
       Builder.getIntN(GRLen, static_cast<uint64_t>(AI->getOrdering()));
   Type *Tys[] = {AlignedAddr->getType()};
   Function *LlwOpScwLoop = Intrinsic::getDeclaration(
       AI->getModule(),
       getIntrinsicForMaskedAtomicRMWBinOp(GRLen, AI->getOperation()), Tys);
 

llvm/test/CodeGen/LoongArch/ir-instruction/atomicrmw.ll

 }
 
 define i8 @atomicrmw_xchg_0_i8_acquire(ptr %a) nounwind {
 ; LA32-LABEL: atomicrmw_xchg_0_i8_acquire:
 ; LA32:       # %bb.0:
 ; LA32-NEXT:    slli.w $a1, $a0, 3
 ; LA32-NEXT:    ori $a2, $zero, 255
 ; LA32-NEXT:    sll.w $a2, $a2, $a1
+; LA32-NEXT:    nor $a2, $a2, $zero
 ; LA32-NEXT:    bstrins.w $a0, $zero, 1, 0
 ; LA32-NEXT:  .LBB1_1: # =>This Inner Loop Header: Depth=1
 ; LA32-NEXT:    dbar 0
 ; LA32-NEXT:    ll.w $a3, $a0, 0
-; LA32-NEXT:    addi.w $a4, $zero, 0
-; LA32-NEXT:    xor $a4, $a3, $a4
-; LA32-NEXT:    and $a4, $a4, $a2
-; LA32-NEXT:    xor $a4, $a3, $a4
+; LA32-NEXT:    and $a4, $a3, $a2
 ; LA32-NEXT:    sc.w $a4, $a0, 0
 ; LA32-NEXT:    beqz $a4, .LBB1_1
 ; LA32-NEXT:  # %bb.2:
 ; LA32-NEXT:    srl.w $a0, $a3, $a1
 ; LA32-NEXT:    ret
 ;
 ; LA64-LABEL: atomicrmw_xchg_0_i8_acquire:
 ; LA64:       # %bb.0:
 ; LA64-NEXT:    slli.d $a1, $a0, 3
 ; LA64-NEXT:    ori $a2, $zero, 255
 ; LA64-NEXT:    sll.w $a2, $a2, $a1
-; LA64-NEXT:    addi.w $a2, $a2, 0
+; LA64-NEXT:    nor $a2, $a2, $zero
 ; LA64-NEXT:    bstrins.d $a0, $zero, 1, 0
-; LA64-NEXT:  .LBB1_1: # =>This Inner Loop Header: Depth=1
-; LA64-NEXT:    dbar 0
-; LA64-NEXT:    ll.w $a3, $a0, 0
-; LA64-NEXT:    addi.w $a4, $zero, 0
-; LA64-NEXT:    xor $a4, $a3, $a4
-; LA64-NEXT:    and $a4, $a4, $a2
-; LA64-NEXT:    xor $a4, $a3, $a4
-; LA64-NEXT:    sc.w $a4, $a0, 0
-; LA64-NEXT:    beqz $a4, .LBB1_1
-; LA64-NEXT:  # %bb.2:
+; LA64-NEXT:    amand_db.w $a3, $a2, $a0
 ; LA64-NEXT:    srl.w $a0, $a3, $a1
 ; LA64-NEXT:    ret
   %1 = atomicrmw xchg ptr %a, i8 0 acquire
   ret i8 %1
 }
 
 define i8 @atomicrmw_xchg_minus_1_i8_acquire(ptr %a) nounwind {
 ; LA32-LABEL: atomicrmw_xchg_minus_1_i8_acquire:
 ; LA32:       # %bb.0:
 ; LA32-NEXT:    slli.w $a1, $a0, 3
 ; LA32-NEXT:    ori $a2, $zero, 255
 ; LA32-NEXT:    sll.w $a2, $a2, $a1
 ; LA32-NEXT:    bstrins.w $a0, $zero, 1, 0
 ; LA32-NEXT:  .LBB2_1: # =>This Inner Loop Header: Depth=1
 ; LA32-NEXT:    dbar 0
 ; LA32-NEXT:    ll.w $a3, $a0, 0
-; LA32-NEXT:    addi.w $a4, $a2, 0
-; LA32-NEXT:    xor $a4, $a3, $a4
-; LA32-NEXT:    and $a4, $a4, $a2
-; LA32-NEXT:    xor $a4, $a3, $a4
+; LA32-NEXT:    or $a4, $a3, $a2
 ; LA32-NEXT:    sc.w $a4, $a0, 0
 ; LA32-NEXT:    beqz $a4, .LBB2_1
 ; LA32-NEXT:  # %bb.2:
 ; LA32-NEXT:    srl.w $a0, $a3, $a1
 ; LA32-NEXT:    ret
 ;
 ; LA64-LABEL: atomicrmw_xchg_minus_1_i8_acquire:
 ; LA64:       # %bb.0:
 ; LA64-NEXT:    slli.d $a1, $a0, 3
 ; LA64-NEXT:    ori $a2, $zero, 255
 ; LA64-NEXT:    sll.w $a2, $a2, $a1
-; LA64-NEXT:    addi.w $a2, $a2, 0
 ; LA64-NEXT:    bstrins.d $a0, $zero, 1, 0
-; LA64-NEXT:  .LBB2_1: # =>This Inner Loop Header: Depth=1
-; LA64-NEXT:    dbar 0
-; LA64-NEXT:    ll.w $a3, $a0, 0
-; LA64-NEXT:    addi.w $a4, $a2, 0
-; LA64-NEXT:    xor $a4, $a3, $a4
-; LA64-NEXT:    and $a4, $a4, $a2
-; LA64-NEXT:    xor $a4, $a3, $a4
-; LA64-NEXT:    sc.w $a4, $a0, 0
-; LA64-NEXT:    beqz $a4, .LBB2_1
-; LA64-NEXT:  # %bb.2:
+; LA64-NEXT:    amor_db.w $a3, $a2, $a0
 ; LA64-NEXT:    srl.w $a0, $a3, $a1
 ; LA64-NEXT:    ret
   %1 = atomicrmw xchg ptr %a, i8 -1 acquire
   ret i8 %1
 }
 
 define i16 @atomicrmw_xchg_i16_acquire(ptr %a, i16 %b) nounwind {
 ; LA32-LABEL: atomicrmw_xchg_i16_acquire:

 
 define i16 @atomicrmw_xchg_0_i16_acquire(ptr %a) nounwind {
 ; LA32-LABEL: atomicrmw_xchg_0_i16_acquire:
 ; LA32:       # %bb.0:
 ; LA32-NEXT:    lu12i.w $a1, 15
 ; LA32-NEXT:    ori $a1, $a1, 4095
 ; LA32-NEXT:    slli.w $a2, $a0, 3
 ; LA32-NEXT:    sll.w $a1, $a1, $a2
+; LA32-NEXT:    nor $a1, $a1, $zero
 ; LA32-NEXT:    bstrins.w $a0, $zero, 1, 0
 ; LA32-NEXT:  .LBB4_1: # =>This Inner Loop Header: Depth=1
 ; LA32-NEXT:    dbar 0
 ; LA32-NEXT:    ll.w $a3, $a0, 0
-; LA32-NEXT:    addi.w $a4, $zero, 0
-; LA32-NEXT:    xor $a4, $a3, $a4
-; LA32-NEXT:    and $a4, $a4, $a1
-; LA32-NEXT:    xor $a4, $a3, $a4
+; LA32-NEXT:    and $a4, $a3, $a1
 ; LA32-NEXT:    sc.w $a4, $a0, 0
 ; LA32-NEXT:    beqz $a4, .LBB4_1
 ; LA32-NEXT:  # %bb.2:
 ; LA32-NEXT:    srl.w $a0, $a3, $a2
 ; LA32-NEXT:    ret
 ;
 ; LA64-LABEL: atomicrmw_xchg_0_i16_acquire:
 ; LA64:       # %bb.0:
 ; LA64-NEXT:    lu12i.w $a1, 15
 ; LA64-NEXT:    ori $a1, $a1, 4095
 ; LA64-NEXT:    slli.d $a2, $a0, 3
 ; LA64-NEXT:    sll.w $a1, $a1, $a2
-; LA64-NEXT:    addi.w $a1, $a1, 0
+; LA64-NEXT:    nor $a1, $a1, $zero
 ; LA64-NEXT:    bstrins.d $a0, $zero, 1, 0
-; LA64-NEXT:  .LBB4_1: # =>This Inner Loop Header: Depth=1
-; LA64-NEXT:    dbar 0
-; LA64-NEXT:    ll.w $a3, $a0, 0
-; LA64-NEXT:    addi.w $a4, $zero, 0
-; LA64-NEXT:    xor $a4, $a3, $a4
-; LA64-NEXT:    and $a4, $a4, $a1
-; LA64-NEXT:    xor $a4, $a3, $a4
-; LA64-NEXT:    sc.w $a4, $a0, 0
-; LA64-NEXT:    beqz $a4, .LBB4_1
-; LA64-NEXT:  # %bb.2:
+; LA64-NEXT:    amand_db.w $a3, $a1, $a0
 ; LA64-NEXT:    srl.w $a0, $a3, $a2
 ; LA64-NEXT:    ret
   %1 = atomicrmw xchg ptr %a, i16 0 acquire
   ret i16 %1
 }
 
 define i16 @atomicrmw_xchg_minus_1_i16_acquire(ptr %a) nounwind {
 ; LA32-LABEL: atomicrmw_xchg_minus_1_i16_acquire:
 ; LA32:       # %bb.0:
 ; LA32-NEXT:    lu12i.w $a1, 15
 ; LA32-NEXT:    ori $a1, $a1, 4095
 ; LA32-NEXT:    slli.w $a2, $a0, 3
 ; LA32-NEXT:    sll.w $a1, $a1, $a2
 ; LA32-NEXT:    bstrins.w $a0, $zero, 1, 0
 ; LA32-NEXT:  .LBB5_1: # =>This Inner Loop Header: Depth=1
 ; LA32-NEXT:    dbar 0
 ; LA32-NEXT:    ll.w $a3, $a0, 0
-; LA32-NEXT:    addi.w $a4, $a1, 0
-; LA32-NEXT:    xor $a4, $a3, $a4
-; LA32-NEXT:    and $a4, $a4, $a1
-; LA32-NEXT:    xor $a4, $a3, $a4
+; LA32-NEXT:    or $a4, $a3, $a1
 ; LA32-NEXT:    sc.w $a4, $a0, 0
 ; LA32-NEXT:    beqz $a4, .LBB5_1
 ; LA32-NEXT:  # %bb.2:
 ; LA32-NEXT:    srl.w $a0, $a3, $a2
 ; LA32-NEXT:    ret
 ;
 ; LA64-LABEL: atomicrmw_xchg_minus_1_i16_acquire:
 ; LA64:       # %bb.0:
 ; LA64-NEXT:    lu12i.w $a1, 15
 ; LA64-NEXT:    ori $a1, $a1, 4095
 ; LA64-NEXT:    slli.d $a2, $a0, 3
 ; LA64-NEXT:    sll.w $a1, $a1, $a2
-; LA64-NEXT:    addi.w $a1, $a1, 0
 ; LA64-NEXT:    bstrins.d $a0, $zero, 1, 0
-; LA64-NEXT:  .LBB5_1: # =>This Inner Loop Header: Depth=1
-; LA64-NEXT:    dbar 0
-; LA64-NEXT:    ll.w $a3, $a0, 0
-; LA64-NEXT:    addi.w $a4, $a1, 0
-; LA64-NEXT:    xor $a4, $a3, $a4
-; LA64-NEXT:    and $a4, $a4, $a1
-; LA64-NEXT:    xor $a4, $a3, $a4
-; LA64-NEXT:    sc.w $a4, $a0, 0
-; LA64-NEXT:    beqz $a4, .LBB5_1
-; LA64-NEXT:  # %bb.2:
+; LA64-NEXT:    amor_db.w $a3, $a1, $a0
 ; LA64-NEXT:    srl.w $a0, $a3, $a2
 ; LA64-NEXT:    ret
   %1 = atomicrmw xchg ptr %a, i16 -1 acquire
   ret i16 %1
 }
 
 define i32 @atomicrmw_xchg_i32_acquire(ptr %a, i32 %b) nounwind {
 ; LA32-LABEL: atomicrmw_xchg_i32_acquire:

 }
 
 define i8 @atomicrmw_xchg_0_i8_monotonic(ptr %a) nounwind {
 ; LA32-LABEL: atomicrmw_xchg_0_i8_monotonic:
 ; LA32:       # %bb.0:
 ; LA32-NEXT:    slli.w $a1, $a0, 3
 ; LA32-NEXT:    ori $a2, $zero, 255
 ; LA32-NEXT:    sll.w $a2, $a2, $a1
+; LA32-NEXT:    nor $a2, $a2, $zero
 ; LA32-NEXT:    bstrins.w $a0, $zero, 1, 0
 ; LA32-NEXT:  .LBB33_1: # =>This Inner Loop Header: Depth=1
 ; LA32-NEXT:    ll.w $a3, $a0, 0
-; LA32-NEXT:    addi.w $a4, $zero, 0
-; LA32-NEXT:    xor $a4, $a3, $a4
-; LA32-NEXT:    and $a4, $a4, $a2
-; LA32-NEXT:    xor $a4, $a3, $a4
+; LA32-NEXT:    and $a4, $a3, $a2
 ; LA32-NEXT:    sc.w $a4, $a0, 0
 ; LA32-NEXT:    beqz $a4, .LBB33_1
 ; LA32-NEXT:  # %bb.2:
 ; LA32-NEXT:    srl.w $a0, $a3, $a1
 ; LA32-NEXT:    ret
 ;
 ; LA64-LABEL: atomicrmw_xchg_0_i8_monotonic:
 ; LA64:       # %bb.0:
 ; LA64-NEXT:    slli.d $a1, $a0, 3
 ; LA64-NEXT:    ori $a2, $zero, 255
 ; LA64-NEXT:    sll.w $a2, $a2, $a1
-; LA64-NEXT:    addi.w $a2, $a2, 0
+; LA64-NEXT:    nor $a2, $a2, $zero
 ; LA64-NEXT:    bstrins.d $a0, $zero, 1, 0
-; LA64-NEXT:  .LBB33_1: # =>This Inner Loop Header: Depth=1
-; LA64-NEXT:    ll.w $a3, $a0, 0
-; LA64-NEXT:    addi.w $a4, $zero, 0
-; LA64-NEXT:    xor $a4, $a3, $a4
-; LA64-NEXT:    and $a4, $a4, $a2
-; LA64-NEXT:    xor $a4, $a3, $a4
-; LA64-NEXT:    sc.w $a4, $a0, 0
-; LA64-NEXT:    beqz $a4, .LBB33_1
-; LA64-NEXT:  # %bb.2:
+; LA64-NEXT:    amand_db.w $a3, $a2, $a0
 ; LA64-NEXT:    srl.w $a0, $a3, $a1
 ; LA64-NEXT:    ret
   %1 = atomicrmw xchg ptr %a, i8 0 monotonic
   ret i8 %1
 }
 
 define i8 @atomicrmw_xchg_minus_1_i8_monotonic(ptr %a) nounwind {
 ; LA32-LABEL: atomicrmw_xchg_minus_1_i8_monotonic:
 ; LA32:       # %bb.0:
 ; LA32-NEXT:    slli.w $a1, $a0, 3
 ; LA32-NEXT:    ori $a2, $zero, 255
 ; LA32-NEXT:    sll.w $a2, $a2, $a1
 ; LA32-NEXT:    bstrins.w $a0, $zero, 1, 0
 ; LA32-NEXT:  .LBB34_1: # =>This Inner Loop Header: Depth=1
 ; LA32-NEXT:    ll.w $a3, $a0, 0
-; LA32-NEXT:    addi.w $a4, $a2, 0
-; LA32-NEXT:    xor $a4, $a3, $a4
-; LA32-NEXT:    and $a4, $a4, $a2
-; LA32-NEXT:    xor $a4, $a3, $a4
+; LA32-NEXT:    or $a4, $a3, $a2
 ; LA32-NEXT:    sc.w $a4, $a0, 0
 ; LA32-NEXT:    beqz $a4, .LBB34_1
 ; LA32-NEXT:  # %bb.2:
 ; LA32-NEXT:    srl.w $a0, $a3, $a1
 ; LA32-NEXT:    ret
 ;
 ; LA64-LABEL: atomicrmw_xchg_minus_1_i8_monotonic:
 ; LA64:       # %bb.0:
 ; LA64-NEXT:    slli.d $a1, $a0, 3
 ; LA64-NEXT:    ori $a2, $zero, 255
 ; LA64-NEXT:    sll.w $a2, $a2, $a1
-; LA64-NEXT:    addi.w $a2, $a2, 0
 ; LA64-NEXT:    bstrins.d $a0, $zero, 1, 0
-; LA64-NEXT:  .LBB34_1: # =>This Inner Loop Header: Depth=1
-; LA64-NEXT:    ll.w $a3, $a0, 0
-; LA64-NEXT:    addi.w $a4, $a2, 0
-; LA64-NEXT:    xor $a4, $a3, $a4
-; LA64-NEXT:    and $a4, $a4, $a2
-; LA64-NEXT:    xor $a4, $a3, $a4
-; LA64-NEXT:    sc.w $a4, $a0, 0
-; LA64-NEXT:    beqz $a4, .LBB34_1
-; LA64-NEXT:  # %bb.2:
+; LA64-NEXT:    amor_db.w $a3, $a2, $a0
 ; LA64-NEXT:    srl.w $a0, $a3, $a1
 ; LA64-NEXT:    ret
   %1 = atomicrmw xchg ptr %a, i8 -1 monotonic
   ret i8 %1
 }
 
 define i16 @atomicrmw_xchg_i16_monotonic(ptr %a, i16 %b) nounwind {
 ; LA32-LABEL: atomicrmw_xchg_i16_monotonic:

 
 define i16 @atomicrmw_xchg_0_i16_monotonic(ptr %a) nounwind {
 ; LA32-LABEL: atomicrmw_xchg_0_i16_monotonic:
 ; LA32:       # %bb.0:
 ; LA32-NEXT:    lu12i.w $a1, 15
 ; LA32-NEXT:    ori $a1, $a1, 4095
 ; LA32-NEXT:    slli.w $a2, $a0, 3
 ; LA32-NEXT:    sll.w $a1, $a1, $a2
+; LA32-NEXT:    nor $a1, $a1, $zero
 ; LA32-NEXT:    bstrins.w $a0, $zero, 1, 0
 ; LA32-NEXT:  .LBB36_1: # =>This Inner Loop Header: Depth=1
 ; LA32-NEXT:    ll.w $a3, $a0, 0
-; LA32-NEXT:    addi.w $a4, $zero, 0
-; LA32-NEXT:    xor $a4, $a3, $a4
-; LA32-NEXT:    and $a4, $a4, $a1
-; LA32-NEXT:    xor $a4, $a3, $a4
+; LA32-NEXT:    and $a4, $a3, $a1
 ; LA32-NEXT:    sc.w $a4, $a0, 0
 ; LA32-NEXT:    beqz $a4, .LBB36_1
 ; LA32-NEXT:  # %bb.2:
 ; LA32-NEXT:    srl.w $a0, $a3, $a2
 ; LA32-NEXT:    ret
 ;
 ; LA64-LABEL: atomicrmw_xchg_0_i16_monotonic:
 ; LA64:       # %bb.0:
 ; LA64-NEXT:    lu12i.w $a1, 15
 ; LA64-NEXT:    ori $a1, $a1, 4095
 ; LA64-NEXT:    slli.d $a2, $a0, 3
 ; LA64-NEXT:    sll.w $a1, $a1, $a2
-; LA64-NEXT:    addi.w $a1, $a1, 0
+; LA64-NEXT:    nor $a1, $a1, $zero
 ; LA64-NEXT:    bstrins.d $a0, $zero, 1, 0
-; LA64-NEXT:  .LBB36_1: # =>This Inner Loop Header: Depth=1
-; LA64-NEXT:    ll.w $a3, $a0, 0
-; LA64-NEXT:    addi.w $a4, $zero, 0
-; LA64-NEXT:    xor $a4, $a3, $a4
-; LA64-NEXT:    and $a4, $a4, $a1
-; LA64-NEXT:    xor $a4, $a3, $a4
-; LA64-NEXT:    sc.w $a4, $a0, 0
-; LA64-NEXT:    beqz $a4, .LBB36_1
-; LA64-NEXT:  # %bb.2:
+; LA64-NEXT:    amand_db.w $a3, $a1, $a0
 ; LA64-NEXT:    srl.w $a0, $a3, $a2
 ; LA64-NEXT:    ret
   %1 = atomicrmw xchg ptr %a, i16 0 monotonic
   ret i16 %1
 }
 
 define i16 @atomicrmw_xchg_minus_1_i16_monotonic(ptr %a) nounwind {
 ; LA32-LABEL: atomicrmw_xchg_minus_1_i16_monotonic:
 ; LA32:       # %bb.0:
 ; LA32-NEXT:    lu12i.w $a1, 15
 ; LA32-NEXT:    ori $a1, $a1, 4095
 ; LA32-NEXT:    slli.w $a2, $a0, 3
 ; LA32-NEXT:    sll.w $a1, $a1, $a2
 ; LA32-NEXT:    bstrins.w $a0, $zero, 1, 0
 ; LA32-NEXT:  .LBB37_1: # =>This Inner Loop Header: Depth=1
 ; LA32-NEXT:    ll.w $a3, $a0, 0
-; LA32-NEXT:    addi.w $a4, $a1, 0
-; LA32-NEXT:    xor $a4, $a3, $a4
-; LA32-NEXT:    and $a4, $a4, $a1
-; LA32-NEXT:    xor $a4, $a3, $a4
+; LA32-NEXT:    or $a4, $a3, $a1
 ; LA32-NEXT:    sc.w $a4, $a0, 0
 ; LA32-NEXT:    beqz $a4, .LBB37_1
 ; LA32-NEXT:  # %bb.2:
 ; LA32-NEXT:    srl.w $a0, $a3, $a2
 ; LA32-NEXT:    ret
 ;
 ; LA64-LABEL: atomicrmw_xchg_minus_1_i16_monotonic:
 ; LA64:       # %bb.0:
 ; LA64-NEXT:    lu12i.w $a1, 15
 ; LA64-NEXT:    ori $a1, $a1, 4095
 ; LA64-NEXT:    slli.d $a2, $a0, 3
 ; LA64-NEXT:    sll.w $a1, $a1, $a2
-; LA64-NEXT:    addi.w $a1, $a1, 0
 ; LA64-NEXT:    bstrins.d $a0, $zero, 1, 0
-; LA64-NEXT:  .LBB37_1: # =>This Inner Loop Header: Depth=1
-; LA64-NEXT:    ll.w $a3, $a0, 0
-; LA64-NEXT:    addi.w $a4, $a1, 0
-; LA64-NEXT:    xor $a4, $a3, $a4
-; LA64-NEXT:    and $a4, $a4, $a1
-; LA64-NEXT:    xor $a4, $a3, $a4
-; LA64-NEXT:    sc.w $a4, $a0, 0
-; LA64-NEXT:    beqz $a4, .LBB37_1
-; LA64-NEXT:  # %bb.2:
+; LA64-NEXT:    amor_db.w $a3, $a1, $a0
 ; LA64-NEXT:    srl.w $a0, $a3, $a2
 ; LA64-NEXT:    ret
   %1 = atomicrmw xchg ptr %a, i16 -1 monotonic
   ret i16 %1
 }
 
 define i32 @atomicrmw_xchg_i32_monotonic(ptr %a, i32 %b) nounwind {
 ; LA32-LABEL: atomicrmw_xchg_i32_monotonic: