[AMDGPU] Handle direct loads to LDS in memory model

Add additional waitcnt insertion to ensure proper ordering between LDS
operations and direct loads from global memory to LDS on pre-GFX10
hardware.

Direct LDS loads perform both a global memory load and an LDS store,
which can be reordered with respect to other LDS operations without
explicit synchronization. This can cause ordering violations even within
a single thread.

The change conservatively inserts vmcnt(0) waits for all sync scopes
when the LDS address space is involved. Future optimizations in
SIInsertWaitcnts can relax this to only wait for outstanding direct LDS
loads rather than all vmcnt events.

This change only affects LDS address space synchronization and preserves
existing cross-address space ordering behavior.

Author: kerbowa

llvm/lib/Target/AMDGPU/SIMemoryLegalizer.cpp

@@ -1084,6 +1084,7 @@ bool SIGfx6CacheControl::insertWait(MachineBasicBlock::iterator &MI,
 
   bool VMCnt = false;
   bool LGKMCnt = false;
+  bool DirectLDSWait = false;
 
   if ((AddrSpace & (SIAtomicAddrSpace::GLOBAL | SIAtomicAddrSpace::SCRATCH)) !=
       SIAtomicAddrSpace::NONE) {
@@ -1104,6 +1105,10 @@ bool SIGfx6CacheControl::insertWait(MachineBasicBlock::iterator &MI,
   }
 
   if ((AddrSpace & SIAtomicAddrSpace::LDS) != SIAtomicAddrSpace::NONE) {
+    // Wait for direct loads to LDS from global memory to ensure that
+    // LDS operations cannot be reordered with respect to global memory
+    // operations.
+    DirectLDSWait = true;
     switch (Scope) {
     case SIAtomicScope::SYSTEM:
     case SIAtomicScope::AGENT:
@@ -1149,6 +1154,18 @@ bool SIGfx6CacheControl::insertWait(MachineBasicBlock::iterator &MI,
     }
   }
 
+  // Conservatively wait for vmcnt(0) to ensure that LDS operations and direct
+  // LDS loads from global memory cannot be reordered with respect to each
+  // other. This waitcnt can be safely optimized to wait for a higher vmcnt
+  // based on the number of outstanding direct LDS loads.
+  if (DirectLDSWait) {
+    unsigned WaitCntImmediate = AMDGPU::encodeWaitcnt(
+        IV, 0, getExpcntBitMask(IV), getLgkmcntBitMask(IV));
+    BuildMI(MBB, MI, DL, TII->get(AMDGPU::S_WAITCNT_DIRECT_LDS_LOAD_soft))
+        .addImm(WaitCntImmediate);
+    Changed = true;
+  }
+
   if (VMCnt || LGKMCnt) {
     unsigned WaitCntImmediate =
       AMDGPU::encodeWaitcnt(IV,