[flang] [clang] [clang-tools-extra] [compiler-rt] [libcxx] [llvm] [libc] [AMDGPU] Fix lack of LDS DMA check in the AA handling (PR #75249)

[Stanislav Mekhanoshin via cfe-commits]

================
@@ -3656,8 +3656,8 @@ bool SIInstrInfo::areMemAccessesTriviallyDisjoint(const MachineInstr &MIa,
   // underlying address space, even if it was lowered to a different one,
   // e.g. private accesses lowered to use MUBUF instructions on a scratch
   // buffer.
-  if (isDS(MIa)) {
-    if (isDS(MIb))
+  if (isDS(MIa) || isLDSDMA(MIa)) {
+    if (isDS(MIb) || isLDSDMA(MIb))
       return checkInstOffsetsDoNotOverlap(MIa, MIb);
----------------
rampitec wrote:

It does, even though it just bails. It goes down to getMemOperandsWithOffsetWidth and there it bails on the LDS DMA:
```
 449│     // Get appropriate operand, and compute width accordingly.
 450│     DataOpIdx = AMDGPU::getNamedOperandIdx(Opc, AMDGPU::OpName::vdst);
 451│     if (DataOpIdx == -1)
 452│       DataOpIdx = AMDGPU::getNamedOperandIdx(Opc, AMDGPU::OpName::vdata);
 453│     if (DataOpIdx == -1) // LDS DMA
 454│       return false;
```
In principle these offsets are analyzable. This is a typical store memop:
```
(dereferenceable store (s32) into `ptr addrspace(3) getelementptr inbounds (%llvm.amdgcn.kernel.buffer_load_lds_dword_2_ar
rays.lds.t, ptr addrspace(3) @llvm.amdgcn.kernel.buffer_load_lds_dword_2_arrays.lds, i32 0, i32 1)
```
But is you want I can bail right here.

https://github.com/llvm/llvm-project/pull/75249