AMDGPU: Report unaligned scratch access as fast if supported by tgt

This enables more consecutive load folding during aggressive-instcombine.

Author: macurtis-amd

llvm/lib/Target/AMDGPU/SIISelLowering.cpp

@@ -2098,10 +2098,16 @@ bool SITargetLowering::allowsMisalignedMemoryAccessesImpl(
   if (AddrSpace == AMDGPUAS::PRIVATE_ADDRESS ||
       AddrSpace == AMDGPUAS::FLAT_ADDRESS) {
     bool AlignedBy4 = Alignment >= Align(4);
+    if (Subtarget->hasUnalignedScratchAccessEnabled()) {
+      if (IsFast)
+        *IsFast = AlignedBy4 ? Size : 1;
+      return true;
+    }
+
     if (IsFast)
       *IsFast = AlignedBy4;
 
-    return AlignedBy4 || Subtarget->hasUnalignedScratchAccessEnabled();
+    return AlignedBy4;
   }
 
   // So long as they are correct, wide global memory operations perform better