[llvm] update P7 v4i8 load cost (PR #108261)

[via llvm-commits]

================
@@ -802,12 +802,18 @@ InstructionCost PPCTTIImpl::getMemoryOpCost(unsigned Opcode, Type *Src,
   // explicitly check this case. There are also corresponding store
   // instructions.
   unsigned MemBytes = Src->getPrimitiveSizeInBits();
-  if (ST->hasVSX() && IsAltivecType &&
-      (MemBytes == 64 || (ST->hasP8Vector() && MemBytes == 32)))
-    return 1;
+  unsigned SrcBytes = LT.second.getStoreSize();
+  if (ST->hasVSX() && IsAltivecType) {
+    if (MemBytes == 64 || (ST->hasP8Vector() && MemBytes == 32))
+      return 1;
+
+    // Use lfiwax/xxspltw
+    Align AlignBytes = Alignment ? *Alignment : Align(1);
+    if (Opcode == Instruction::Load && MemBytes == 32 && AlignBytes < SrcBytes)
----------------
RolandF77 wrote:

If a partial vector (< 128 bits) is being loaded with a full vector aligned address (>= 128 bits), the load will be done as a full vector load since we know from alignment that it is safe. Therefore the cost of a partial vector load does
not apply.

https://github.com/llvm/llvm-project/pull/108261