[PATCH] D25551: AMDGPU: Implement SGPR spilling with scalar stores

[Nicolai Hähnle via llvm-commits]

nhaehnle added inline comments.


================
Comment at: lib/Target/AMDGPU/SIInsertWaits.cpp:626-627
+
+      if (I->getOpcode() == AMDGPU::S_ENDPGM)
+        EndPgmBlocks.push_back(&MBB);
     }
----------------
arsenm wrote:
> nhaehnle wrote:
> > Please also handle the SI_RETURN case here and below.
> I was specifically not handling that, but I guess it isn't a normal function return
Yes. The way it's used, we're just concatenating the binaries of multiple shader parts. Only the middle part should ever contain register spills (well, unless you compile with -O0, but we never do that), so it makes sense that all the necessary handling is confined to the shader part where it happens.


https://reviews.llvm.org/D25551