[PATCH] D98515: [AMDGPU][GlobalISel] Stop foldInsertEltToCmpSelect from changing reg banks

Matt Arsenault via Phabricator via llvm-commits llvm-commits at lists.llvm.org
Thu May 6 18:05:35 PDT 2021


arsenm requested changes to this revision.
arsenm added inline comments.
This revision now requires changes to proceed.


================
Comment at: llvm/lib/Target/AMDGPU/AMDGPURegisterBankInfo.cpp:2002
+    Register InReg = MI.getOperand(2).getReg();
+    if (DstBank == AMDGPU::VGPRRegBank && InsBank == AMDGPU::SGPRRegBank)
+      InReg = B.buildCopy(MRI.getType(InReg), InReg).getReg(0);
----------------
OK I see what's happening now. I think this should introduce a new utility function to try to constrain a register to a register bank, or insert a copy if not (similar to the existing ones for concrete register classes). This should also be called below in place of the setRegBank where this is overwritten


CHANGES SINCE LAST ACTION
  https://reviews.llvm.org/D98515/new/

https://reviews.llvm.org/D98515



More information about the llvm-commits mailing list