commit ebc17a0d68208a967fe8e13e1874874228dda622
author Matt Arsenault <Matthew.Arsenault@amd.com> Tue Jul 27 11:08:06 2021 -0400
committer Matt Arsenault <Matthew.Arsenault@amd.com> Sat Jul 31 10:37:15 2021 -0400
tree 3de923f06c5c4f0920950b4c9d8c45b577dd888d
parent 43c7cb9a3cf528a6e4e81acb6752b273c6e60300

GlobalISel: Scalarize unaligned vector stores

This has the same problems and limitations as the load path.

llvm/lib/CodeGen/GlobalISel/LegalizerHelper.cpp

diff --git a/llvm/lib/CodeGen/GlobalISel/LegalizerHelper.cpp b/llvm/lib/CodeGen/GlobalISel/LegalizerHelper.cpp
index 843d6ae..51d7bc6 100644
--- a/llvm/lib/CodeGen/GlobalISel/LegalizerHelper.cpp
+++ b/llvm/lib/CodeGen/GlobalISel/LegalizerHelper.cpp
@@ -3029,13 +3029,13 @@
   MachineMemOperand &MMO = **StoreMI.memoperands_begin();
   LLT MemTy = MMO.getMemoryType();
 
-  if (SrcTy.isVector())
-    return UnableToLegalize;
-
   unsigned StoreWidth = MemTy.getSizeInBits();
   unsigned StoreSizeInBits = 8 * MemTy.getSizeInBytes();
 
   if (StoreWidth != StoreSizeInBits) {
+    if (SrcTy.isVector())
+      return UnableToLegalize;
+
     // Promote to a byte-sized store with upper bits zero if not
     // storing an integral number of bytes.  For example, promote
     // TRUNCSTORE:i1 X -> TRUNCSTORE:i8 (and X, 1)
@@ -3056,6 +3056,16 @@
     return Legalized;
   }
 
+  if (MemTy.isVector()) {
+    // TODO: Handle vector trunc stores
+    if (MemTy != SrcTy)
+      return UnableToLegalize;
+
+    // TODO: We can do better than scalarizing the vector and at least split it
+    // in half.
+    return reduceLoadStoreWidth(StoreMI, 0, SrcTy.getElementType());
+  }
+
   unsigned MemSizeInBits = MemTy.getSizeInBits();
   uint64_t LargeSplitSize, SmallSplitSize;