[PATCH] D63510: [LiveInterval] Removed bogus empty subrange assert

[Matt Arsenault via Phabricator via llvm-commits]

arsenm added inline comments.


================
Comment at: test/CodeGen/AMDGPU/empty-subrange.mir:1
+# RUN: llc -mtriple=amdgcn -mcpu=gfx802 -verify-machineinstrs -start-after amdgpu-isel -o - %s | FileCheck -check-prefix=GCN %s
+#
----------------
tpr wrote:
> arsenm wrote:
> > Should be restricted to a narrow range of passes in regalloc
> I tried capturing mir just before the live intervals before regalloc, which is the one where it fell over, and running from there, but it did not reproduce the failure. I think it is too sensitive to what order it sees the MOs when iterating, which would be potentially different between reading a mir file and getting real output from earlier passes.
This makes getting this test reduced even more important then. You've dropped all of the machineFunctionInfo, maybe some of those fields help? Spending hours debugging why I couldn't reproduce something like this is why I finally implemented the serialization for it. We don't preserve everything that matters yet


Repository:
  rL LLVM

CHANGES SINCE LAST ACTION
  https://reviews.llvm.org/D63510/new/

https://reviews.llvm.org/D63510