[clang] [llvm] [mlir] [AMDGPU] Add a new amdgcn.load.to.lds intrinsic (PR #137425)

[Krzysztof Drewniak via cfe-commits]

================
@@ -2641,6 +2641,28 @@ def int_amdgcn_perm :
 // GFX9 Intrinsics
 //===----------------------------------------------------------------------===//
 
+/// This is a general-purpose intrinsic for all operations that take a pointer
+/// a base location in LDS, and a data size and use it to perform a gather to LDS.
+/// This allows abstracting over both global pointers (address space 1) and
+/// the buffer-resource-wrapper pointers (address space 7 and 9).
+/// TODO: add support for address space 5 and scratch_load_lds.
+class AMDGPULoadToLDS :
+  Intrinsic <
+    [],
+    [llvm_anyptr_ty,                    // Base pointer to load from. Varies per lane.
+     LLVMQualPointerType<3>,            // LDS base pointer to store to. Must be wave-uniform.
+     llvm_i32_ty,                       // Data byte size: 1/2/4 (/12/16 for gfx950)
+     llvm_i32_ty,                       // imm offset (applied to both input and LDS address)
----------------
krzysz00 wrote:

... Oh. Consider the case that you have global p + N and LDS q + N. Then the LDS combiner can rewrite this to (q' + O) + N, aka q' + (O + N).

Then the two pointers won't have the same offset anymore and so it's unclear if you can slide it onto the instruction immediate

https://github.com/llvm/llvm-project/pull/137425