[AArch64] Teach DAGCombiner that converting two consecutive loads into a vector load is not a good transform when paired loads are available.

The combiner was creating Q-register loads and stores, which then had to be spilled because there are no callee-save Q registers!

llvm-svn: 214634

llvm/lib/CodeGen/SelectionDAG/DAGCombiner.cpp

@@ -9414,6 +9414,13 @@ bool DAGCombiner::MergeConsecutiveStores(StoreSDNode* St) {
   if (LoadNodes.size() < 2)
     return false;
 
+  // If we have load/store pair instructions and we only have two values,
+  // don't bother.
+  unsigned RequiredAlignment;
+  if (LoadNodes.size() == 2 && TLI.hasPairedLoad(MemVT, RequiredAlignment) &&
+      St->getAlignment() >= RequiredAlignment)
+    return false;
+
   // Scan the memory operations on the chain and find the first non-consecutive
   // load memory address. These variables hold the index in the store node
   // array.

llvm/test/CodeGen/AArch64/paired-load.ll

@@ -0,0 +1,16 @@
+; RUN: llc < %s | FileCheck %s
+target datalayout = "e-m:o-i64:64-i128:128-n32:64-S128"
+target triple = "arm64-apple-ios5.0.0"
+
+; Ensure we're generating ldp instructions instead of ldr Q.
+; CHECK: ldp
+; CHECK: stp
+define void @f(i64* %p, i64* %q) {
+  %addr2 = getelementptr i64* %q, i32 1
+  %addr = getelementptr i64* %p, i32 1
+  %x = load i64* %p
+  %y = load i64* %addr
+  store i64 %x, i64* %q
+  store i64 %y, i64* %addr2
+  ret void
+}