[llvm] [AArch64][GlobalISel] Lower G_BUILD_VECTOR to G_INSERT_VECTOR_ELT (PR #105686)

[via llvm-commits]

================
@@ -61,15 +61,41 @@ bool GIMatchTableExecutor::isBaseWithConstantOffset(
 
 bool GIMatchTableExecutor::isObviouslySafeToFold(MachineInstr &MI,
                                                  MachineInstr &IntoMI) const {
+  auto IntoMIIter = IntoMI.getIterator();
+
   // Immediate neighbours are already folded.
   if (MI.getParent() == IntoMI.getParent() &&
-      std::next(MI.getIterator()) == IntoMI.getIterator())
+      std::next(MI.getIterator()) == IntoMIIter)
     return true;
 
   // Convergent instructions cannot be moved in the CFG.
   if (MI.isConvergent() && MI.getParent() != IntoMI.getParent())
     return false;
 
-  return !MI.mayLoadOrStore() && !MI.mayRaiseFPException() &&
-         !MI.hasUnmodeledSideEffects() && MI.implicit_operands().empty();
+  if (MI.isLoadFoldBarrier())
----------------
chuongg3 wrote:

I chose to include the TableGen patch to reduce the amount of regressions this patch creates to help with code review. This lowering patch was intended to help with selecting LD1 instructions but some of the patterns for LD1 were not matching due to reasons decribed in #101675.

https://github.com/llvm/llvm-project/pull/105686