[PATCH] D75138: [WIP][AMDGPU] Add Scratch Wave Offset to Scratch Buffer Descriptor in entry functions

Mon Mar 9 13:30:51 PDT 2020

arsenm added inline comments.

================
Comment at: llvm/lib/Target/AMDGPU/SIFrameLowering.cpp:290
+  unsigned NumPreloadedSGPRs = MFI->getNumPreloadedSGPRs();
+  // FIXME: This is just lifted from AMDGPUAsmPrinter, because I'm not
+  // sure where/if we track InReg SGPR arguments otherwise.
----------------
scott.linder wrote:
> arsenm wrote:
> > This should not need to inspect the original IR. Why can't this just read it directly from MFI? They should be accounted there already?
> @arsenm @nhaehnle I don't think I understand how `inreg` currently works relative to "preloaded" SGPRs; is/should `inreg` be recorded somewhere in the machine function info so this isn't necessary?
Not directly. There shouldn't be any repeating of the calling convention logic here. Either the number of SGPR arguments should be recorded, or it should be inferred from the machine code.

It might be correct to just count the number of SGPR in the function live-in list. I think live in registers can be deleted from the list if they are proven to be unused, so this might be fragile. Finding the highest live in SGPR number may also work.

Repository:
  rG LLVM Github Monorepo

CHANGES SINCE LAST ACTION
  https://reviews.llvm.org/D75138/new/

https://reviews.llvm.org/D75138