[PATCH] D60135: [Pipeliner] Incorrect loop carried dependence calculation

[Denis Antrushin via Phabricator via llvm-commits]

dantrushin created this revision.
dantrushin added a reviewer: bcahoon.
Herald added a subscriber: hiraditya.
Herald added a project: LLVM.

SwingSchedulerDAG::isLoopCarriedDep performs incorrect dependence
calculations. In particular, it cannot properly handle negative
offsets like this:

  %1 = PHI %0, %2
  %x = LD [%1, -6]   // 32-bit load
  ...
  ST %y, [%1, 2]      // 32-bit store
  %2 = ADD %1, 4

and case of stride smaller than access size like this:

  %1 = PHI %0, %2
  %x = LD [%1, 0]        // 32-bit load
  ...
  ST %y, [%1, 2]        // 32-bit store
  %2 = ADD %1, 2


Repository:
  rG LLVM Github Monorepo

https://reviews.llvm.org/D60135

Files:
  llvm/lib/CodeGen/MachinePipeliner.cpp


Index: llvm/lib/CodeGen/MachinePipeliner.cpp
===================================================================
--- llvm/lib/CodeGen/MachinePipeliner.cpp
+++ llvm/lib/CodeGen/MachinePipeliner.cpp
@@ -3167,12 +3167,14 @@
 
   // This is the main test, which checks the offset values and the loop
   // increment value to determine if the accesses may be loop carried.
-  if (OffsetS >= OffsetD)
-    return OffsetS + AccessSizeS > DeltaS;
-  else
-    return OffsetD + AccessSizeD > DeltaD;
+  if (AccessSizeS == MemoryLocation::UnknownSize ||
+      AccessSizeD == MemoryLocation::UnknownSize)
+    return true;
 
-  return true;
+  if (DeltaS != DeltaD || DeltaS < AccessSizeS || DeltaD < AccessSizeD)
+    return true;
+
+  return (OffsetS + (int64_t)AccessSizeS < OffsetD + (int64_t)AccessSizeD);
 }
 
 void SwingSchedulerDAG::postprocessDAG() {


-------------- next part --------------
A non-text attachment was scrubbed...
Name: D60135.193309.patch
Type: text/x-patch
Size: 850 bytes
Desc: not available
URL: <http://lists.llvm.org/pipermail/llvm-commits/attachments/20190402/b18ad332/attachment.bin>