[llvm-commits] [PATCH] [msan] Heuristic handling on unknown intrinsics

[Evgeniy Stepanov]

================
Comment at: llvm/lib/Transforms/Instrumentation/MemorySanitizer.cpp:1240
@@ +1239,3 @@
+        } else {
+          // Two pointer operands? Meh.
+          visitInstruction(I);
----------------
Kostya Serebryany wrote:
> Does this ever happen?
> Maybe add an assert for debug mode? 
> In release more we'll call visitInstruction
It does.
gcread, gcwrite, init_trampoline, adjust_trampoline, neon_vld2lane and friends (ARM).
We either don't want to instrument these, or want to special-case them.

================
Comment at: llvm/lib/Transforms/Instrumentation/MemorySanitizer.cpp:1249
@@ +1248,3 @@
+
+    if (pointerOpIdx < 0 && (readsMemory || writesMemory)) {
+      // No idea how to handle this.
----------------
Kostya Serebryany wrote:
> same here. Maybe add an assert? 
It happens, for example, with memory barriers.

================
Comment at: llvm/lib/Transforms/Instrumentation/MemorySanitizer.cpp:1331
@@ +1330,3 @@
+      Value *ShadowPtr = getShadowPtr(Op, MemAccessShadowTy, IRB);
+      IRB.CreateAlignedStore(Shadow, ShadowPtr, 1);
+      if (ClTrackOrigins)
----------------
Kostya Serebryany wrote:
> This 'alignment 1' deserves a comment
Done. There is a related comment about the unaligned load above, btw.


http://llvm-reviews.chandlerc.com/D184