[llvm] [AArch64][GlobalISel] Adopt some Ld1Lane* patterns for GlobalISel to reduce codegen regressions (PR #69607)

[Vladislav Dzhidzhoev via llvm-commits]

================
@@ -62,9 +63,16 @@ bool GIMatchTableExecutor::isBaseWithConstantOffset(
 bool GIMatchTableExecutor::isObviouslySafeToFold(MachineInstr &MI,
                                                  MachineInstr &IntoMI) const {
   // Immediate neighbours are already folded.
-  if (MI.getParent() == IntoMI.getParent() &&
-      std::next(MI.getIterator()) == IntoMI.getIterator())
-    return true;
+  // Any G_CONSTANT between immediate neighbours can be ignored.
+  if (MI.getParent() == IntoMI.getParent()) {
+    auto IntoIt = IntoMI.getIterator();
+    auto NextIt = std::next(MI.getIterator());
+    while (!NextIt.isEnd() && NextIt != IntoIt &&
+           NextIt->getOpcode() == TargetOpcode::G_CONSTANT)
+      ++NextIt;
+    if (NextIt == IntoIt)
+      return true;
+  }
----------------
dzhidzhoev wrote:

> I feel like there's a more general solution somewhere but I can't think of it at the moment.

Do you consider this credible? https://github.com/llvm/llvm-project/pull/70830

https://github.com/llvm/llvm-project/pull/69607