[llvm-commits] [llvm] r168887 - /llvm/trunk/lib/Transforms/Instrumentation/MemorySanitizer.cpp

Evgeniy Stepanov eugeni.stepanov at gmail.com
Thu Nov 29 06:44:00 PST 2012


Author: eugenis
Date: Thu Nov 29 08:44:00 2012
New Revision: 168887

URL: http://llvm.org/viewvc/llvm-project?rev=168887&view=rev
Log:
[msan] Fix getOriginForNaryOp.
    
The old version failed on a 3-arg instruction with (-1, 0, 0) shadows (it would
pick the 3rd operand origin irrespective of its shadow).
    
The new version always picks the origin of the rightmost poisoned operand.


Modified:
    llvm/trunk/lib/Transforms/Instrumentation/MemorySanitizer.cpp

Modified: llvm/trunk/lib/Transforms/Instrumentation/MemorySanitizer.cpp
URL: http://llvm.org/viewvc/llvm-project/llvm/trunk/lib/Transforms/Instrumentation/MemorySanitizer.cpp?rev=168887&r1=168886&r2=168887&view=diff
==============================================================================
--- llvm/trunk/lib/Transforms/Instrumentation/MemorySanitizer.cpp (original)
+++ llvm/trunk/lib/Transforms/Instrumentation/MemorySanitizer.cpp Thu Nov 29 08:44:00 2012
@@ -842,15 +842,16 @@
   ///
   /// This is a general case of origin propagation. For an Nary operation,
   /// is set to the origin of an argument that is not entirely initialized.
+  /// If there is more than one such arguments, the rightmost of them is picked.
   /// It does not matter which one is picked if all arguments are initialized.
   void setOriginForNaryOp(Instruction &I) {
     if (!ClTrackOrigins) return;
     IRBuilder<> IRB(&I);
     Value *Origin = getOrigin(&I, 0);
     for (unsigned Op = 1, n = I.getNumOperands(); Op < n; ++Op) {
-      Value *S = convertToShadowTyNoVec(getShadow(&I, Op - 1), IRB);
+      Value *S = convertToShadowTyNoVec(getShadow(&I, Op), IRB);
       Origin = IRB.CreateSelect(IRB.CreateICmpNE(S, getCleanShadow(S)),
-                                Origin, getOrigin(&I, Op));
+                                getOrigin(&I, Op), Origin);
     }
     setOrigin(&I, Origin);
   }





More information about the llvm-commits mailing list