[PATCH] D88823: [CodeGen][TailDuplicator] Don't duplicate blocks with INLINEASM_BR

[Bill Wendling via Phabricator via llvm-commits]

void updated this revision to Diff 296216.
void added a comment.

Improve comment on why we're limiting tail duplication on INLINEASM_BR instruction.


Repository:
  rG LLVM Github Monorepo

CHANGES SINCE LAST ACTION
  https://reviews.llvm.org/D88823/new/

https://reviews.llvm.org/D88823

Files:
  llvm/lib/CodeGen/TailDuplicator.cpp


Index: llvm/lib/CodeGen/TailDuplicator.cpp
===================================================================
--- llvm/lib/CodeGen/TailDuplicator.cpp
+++ llvm/lib/CodeGen/TailDuplicator.cpp
@@ -627,6 +627,14 @@
     if (PreRegAlloc && MI.isCall())
       return false;
 
+    // Copying a block with an INLINEASM_BR instruction may result in a PHI node
+    // on the indirect path. We assume that all values used on the indirect path
+    // dominates all paths into the indirect block, i.e. don't have PHI nodes.
+    // FIXME: This may be too restrictive. Perhaps should restrict only if a
+    //        value in the current block isn't used on the indirect path.
+    if (MI.getOpcode() == TargetOpcode::INLINEASM_BR)
+      return false;
+
     if (MI.isBundle())
       InstrCount += MI.getBundleSize();
     else if (!MI.isPHI() && !MI.isMetaInstruction())


-------------- next part --------------
A non-text attachment was scrubbed...
Name: D88823.296216.patch
Type: text/x-patch
Size: 866 bytes
Desc: not available
URL: <http://lists.llvm.org/pipermail/llvm-commits/attachments/20201005/5c9a2ab6/attachment.bin>