[JITLink][MachO][arm64] Mask high bits out of immediate for LDRLiteral19.

Negative deltas for LDRLiteral19 have their high bits set. If these bits aren't
masked out then they will overwrite other instruction bits, leading to a bogus
encoding.

This long-standing relocation bug was exposed by e50aea58d59, "[JITLink][ORC]
Major JITLinkMemoryManager refactor.", which caused memory layouts to be
reordered, which in turn lead to a previously unseen negative delta. (Unseen
because LDRLiteral19s were only created in JITLink passes where they always
pointed at segments that were layed-out-after in the old layout).

No testcase yet: Our existing regression test infrastructure is good at checking
that operand bits are correct, but provides no easy way to test for bad opcode
bits. I'll have a think about the right way to approach this.

https://llvm.org/PR52153

diff --git a/llvm/lib/ExecutionEngine/JITLink/MachO_arm64.cpp b/llvm/lib/ExecutionEngine/JITLink/MachO_arm64.cpp
index bf81fc4f8db9c285ba477e050ea98af93e8cfaf8..af1d88333c85912f592ee756e763dacea8f2e909 100644 (file)
--- a/llvm/lib/ExecutionEngine/JITLink/MachO_arm64.cpp
+++ b/llvm/lib/ExecutionEngine/JITLink/MachO_arm64.cpp
@@ -627,7 +627,8 @@ private:
       if (Delta < -(1 << 20) || Delta > ((1 << 20) - 1))
         return makeTargetOutOfRangeError(G, B, E);
 
-      uint32_t EncodedImm = (static_cast<uint32_t>(Delta) >> 2) << 5;
+      uint32_t EncodedImm =
+        ((static_cast<uint32_t>(Delta) >> 2) & 0x7ffff) << 5;
       uint32_t FixedInstr = RawInstr | EncodedImm;
       *(ulittle32_t *)FixupPtr = FixedInstr;
       break;