[PATCH] D148824: AMDGPU: Define sub-class of SGPR_64 for tail call return

Wed Apr 26 14:49:54 PDT 2023

cfang added inline comments.

================
Comment at: llvm/lib/Target/AMDGPU/SIRegisterInfo.td:830
+def Gfx_CCR_SGPR_64 : SIRegisterClass<"AMDGPU", SGPR_64.RegTypes, 32,
+                                (add (trunc (shl SGPR_64, 15), 1), // s[30:31]
+                                     (trunc (shl SGPR_64, 18), 14))> { // s[36:37]-s[s62:63]
----------------
cdevadas wrote:
> cfang wrote:
> > sebastian-ne wrote:
> > > cdevadas wrote:
> > > > @sebastian-ne can you review this part, the gfx scratch SGPRs?
> > > > 
> > > > The rest of the patch LGTM.
> > > Thanks for the fix!
> > > 
> > > The `s[36:37]-s[s62:63]` range looks good. Why is `s[30:31]` part of `Gfx_CCR_SGPR_64` when it is not part of `CCR_SGPR_64`?
> > Actually I just try to recover what were defined before. My understanding is that s[30:31] is still CSR and we want to avoid CSR completely.  I am not sure about  Gfx_CCR_SGPR_64.
> > 
> > Matt and CD, can you confirm that s[30:31] should not be used as the return address of tail call?  
> I think for the compute side we should not use S[30:31] for tail call target addresses as they are now CSRs, and their spills/restores at prolog/epilog are needed for the debug unwinding info. 
> 
> def CCR_SGPR_64 : SIRegisterClass<"AMDGPU", SGPR_64.RegTypes, 32, (add (trunc SGPR_64, **16**))
> 
> Here S[30:31] pair is currently part of CCR_SGPR_64. The immediate operand in trunc should be 15 instead of 16 to exclude it.
Should we should also exclude s[30:31] for Gfx_CCR_SGPR_64. What do you think, @sebastian-ne ?

CHANGES SINCE LAST ACTION
  https://reviews.llvm.org/D148824/new/

https://reviews.llvm.org/D148824