[llvm] [AMDGPU] Add writelane and readlane pseudos for SGPR spilling (PR #69923)

Carl Ritson via llvm-commits llvm-commits at lists.llvm.org
Tue Oct 24 20:50:40 PDT 2023


================
@@ -3907,7 +3915,9 @@ bool SIInstrInfo::hasUnwantedEffectsWhenEXECEmpty(const MachineInstr &MI) const
   // However, executing them with EXEC = 0 causes them to operate on undefined
----------------
perlfu wrote:

I notice this comment is only really accurate for V_READFIRSTLANE which depends on EXEC.
readlane and writelane are well defined with EXEC=0 -- just we have other reasons for not running them.
If we are sure we want to include spill/reload here then they should probably be documented with a separate comment, because really we are avoiding these in EXEC=0 because it means a code branch should be run for *any* scalar operation.

https://github.com/llvm/llvm-project/pull/69923


More information about the llvm-commits mailing list