[PATCH] D107534: Add a pass to garbage-collect empty basic blocks after code generation.

[David Blaikie via Phabricator via llvm-commits]

dblaikie added a subscriber: jmorse.
dblaikie added inline comments.


================
Comment at: llvm/lib/CodeGen/GCEmptyBasicBlocks.cpp:53
+    if (HasAnyRealCode)
+      continue;
+
----------------
rahmanl wrote:
> wenlei wrote:
> > rahmanl wrote:
> > > wenlei wrote:
> > > > tmsriram wrote:
> > > > > If the block has no real instructions but debug instructions, is it correct to nuke it?
> > > > For debuggbility, probably not. Relatedly the current implementation also keep the block if it contains pseudo-probe, which is what we want. 
> > > I am guessing debug-info instructions in a basic block with zero binary size is never used. Any ideas @dblaikie ?
> > > 
> > > @wenlei My initial idea was to GC all empty blocks before inserting pseudo probes. Now I realize pseudo probes may be inserted anytime. But I still think we can remove empty blocks with pseudo probes at this point of CodeGen. WDYT?
> > pseudo-probes are inserted very early - similar to where IRPGO probes are inserted. If a block has pseudo-probe in it, we should not remove that block, because the purpose of pseudo-probe is to preserve control flow (branches) so profile can be mapped to unoptimized IR more accurately. 
> Thanks for the explanation. I got a bit confused looking at where PseudoProbeInserter is added in TargetPassConfig.cpp, which is only for callsites.
@jmorse for debug location intrinsic functions - but my guess would be deleting an unreachable block would be fine. I'd guess SimplifyCFG already does this? (you could try it -and if it does just delete the block, despite the presence of debug location intrinsics, it's probably a good sign you can do that too)


Repository:
  rG LLVM Github Monorepo

CHANGES SINCE LAST ACTION
  https://reviews.llvm.org/D107534/new/

https://reviews.llvm.org/D107534