[PATCH] D83394: [AMDGPU] Avoid splitting FLAT offsets in unsafe ways

[Nicolai Hähnle via Phabricator via llvm-commits]

nhaehnle added inline comments.


================
Comment at: llvm/lib/Target/AMDGPU/AMDGPUISelDAGToDAG.cpp:1698
+        // into two pieces that are both >= 0 or both <= 0.
 
         SDLoc DL(N);
----------------
arsenm wrote:
> Can you add a fixme to check if this is needed if we know the address isn't FLAT_ADDRESS?
Yes, I'm pretty sure it's needed. If we generate a FLAT instruction, the hardware will check apertures regardless, and you could run into the following scenario:
```
  end_of_scratch_aperture: 0x2'0000'0000
  vaddr: 0x1'ffff'fff0
  inst_offset: 32
```
The hardware executes the flat instruction, sees that vaddr falls into the scratch aperture, and executes the instruction as accessing scratch memory.

The only way to fix that would be if we knew that there is never any memory mapped directly after the end of any of the apertures.


Repository:
  rG LLVM Github Monorepo

CHANGES SINCE LAST ACTION
  https://reviews.llvm.org/D83394/new/

https://reviews.llvm.org/D83394