[llvm] [AMDGPU] Improve isBasicBlockPrologue helper function (PR #69924)

Christudasan Devadasan via llvm-commits llvm-commits at lists.llvm.org
Thu Nov 9 07:42:16 PST 2023


cdevadas wrote:

> Ping. Is this being looked into? We have reverted the change in the graphics downstream branch, but the issue exists in trunk.

I'm looking into it. The assertion started coming because the insertion point is now modified. SplitKit expects the COPY during liverange split to be inserted at the top of the BB. But after my patch, the insertion point vary and they might not be always inserted before the first instruction. One or the other way, the greedy and Fastalloc have problem with the recomputed bb-prolog. I'm trying to see if the insertion point can be retained for SGPR spills/copies as an immediate fixup. But isBasicBlockPrologue being a common interface used for InlineSpiller, SplitKit and MachineIRSink to identify the insertion point. It is becoming difficult to handle the case.  I will provide an update soon.

https://github.com/llvm/llvm-project/pull/69924


More information about the llvm-commits mailing list