[llvm] [AMDGPU][LSV] Restrict large vectors (PR #92540)

[Matt Arsenault via llvm-commits]

================
@@ -364,19 +364,9 @@ unsigned GCNTTIImpl::getStoreVectorFactor(unsigned VF, unsigned StoreSize,
 }
 
 unsigned GCNTTIImpl::getLoadStoreVecRegBitWidth(unsigned AddrSpace) const {
-  if (AddrSpace == AMDGPUAS::GLOBAL_ADDRESS ||
-      AddrSpace == AMDGPUAS::CONSTANT_ADDRESS ||
-      AddrSpace == AMDGPUAS::CONSTANT_ADDRESS_32BIT ||
-      AddrSpace == AMDGPUAS::BUFFER_FAT_POINTER ||
-      AddrSpace == AMDGPUAS::BUFFER_RESOURCE ||
-      AddrSpace == AMDGPUAS::BUFFER_STRIDED_POINTER) {
-    return 512;
-  }
-
   if (AddrSpace == AMDGPUAS::PRIVATE_ADDRESS)
     return 8 * ST->getMaxPrivateElementSize();
----------------
arsenm wrote:

Separate patch, but this is wrong when using scratch instructions. It should be the same as any other flat / global 

https://github.com/llvm/llvm-project/pull/92540