Home | History | Annotate | Line # | Download | only in xray
      1 //===-- xray_AArch64.cc -----------------------------------------*- C++ -*-===//
      2 //
      3 //                     The LLVM Compiler Infrastructure
      4 //
      5 // This file is distributed under the University of Illinois Open Source
      6 // License. See LICENSE.TXT for details.
      7 //
      8 //===----------------------------------------------------------------------===//
      9 //
     10 // This file is a part of XRay, a dynamic runtime instrumentation system.
     11 //
     12 // Implementation of AArch64-specific routines (64-bit).
     13 //
     14 //===----------------------------------------------------------------------===//
     15 #include "sanitizer_common/sanitizer_common.h"
     16 #include "xray_defs.h"
     17 #include "xray_interface_internal.h"
     18 #include <atomic>
     19 #include <cassert>
     20 
     21 extern "C" void __clear_cache(void *start, void *end);
     22 
     23 namespace __xray {
     24 
     25 // The machine codes for some instructions used in runtime patching.
     26 enum class PatchOpcodes : uint32_t {
     27   PO_StpX0X30SP_m16e = 0xA9BF7BE0, // STP X0, X30, [SP, #-16]!
     28   PO_LdrW0_12 = 0x18000060,        // LDR W0, #12
     29   PO_LdrX16_12 = 0x58000070,       // LDR X16, #12
     30   PO_BlrX16 = 0xD63F0200,          // BLR X16
     31   PO_LdpX0X30SP_16 = 0xA8C17BE0,   // LDP X0, X30, [SP], #16
     32   PO_B32 = 0x14000008              // B #32
     33 };
     34 
     35 inline static bool patchSled(const bool Enable, const uint32_t FuncId,
     36                              const XRaySledEntry &Sled,
     37                              void (*TracingHook)()) XRAY_NEVER_INSTRUMENT {
     38   // When |Enable| == true,
     39   // We replace the following compile-time stub (sled):
     40   //
     41   // xray_sled_n:
     42   //   B #32
     43   //   7 NOPs (24 bytes)
     44   //
     45   // With the following runtime patch:
     46   //
     47   // xray_sled_n:
     48   //   STP X0, X30, [SP, #-16]! ; PUSH {r0, lr}
     49   //   LDR W0, #12 ; W0 := function ID
     50   //   LDR X16,#12 ; X16 := address of the trampoline
     51   //   BLR X16
     52   //   ;DATA: 32 bits of function ID
     53   //   ;DATA: lower 32 bits of the address of the trampoline
     54   //   ;DATA: higher 32 bits of the address of the trampoline
     55   //   LDP X0, X30, [SP], #16 ; POP {r0, lr}
     56   //
     57   // Replacement of the first 4-byte instruction should be the last and atomic
     58   // operation, so that the user code which reaches the sled concurrently
     59   // either jumps over the whole sled, or executes the whole sled when the
     60   // latter is ready.
     61   //
     62   // When |Enable|==false, we set back the first instruction in the sled to be
     63   //   B #32
     64 
     65   uint32_t *FirstAddress = reinterpret_cast<uint32_t *>(Sled.Address);
     66   uint32_t *CurAddress = FirstAddress + 1;
     67   if (Enable) {
     68     *CurAddress = uint32_t(PatchOpcodes::PO_LdrW0_12);
     69     CurAddress++;
     70     *CurAddress = uint32_t(PatchOpcodes::PO_LdrX16_12);
     71     CurAddress++;
     72     *CurAddress = uint32_t(PatchOpcodes::PO_BlrX16);
     73     CurAddress++;
     74     *CurAddress = FuncId;
     75     CurAddress++;
     76     *reinterpret_cast<void (**)()>(CurAddress) = TracingHook;
     77     CurAddress += 2;
     78     *CurAddress = uint32_t(PatchOpcodes::PO_LdpX0X30SP_16);
     79     CurAddress++;
     80     std::atomic_store_explicit(
     81         reinterpret_cast<std::atomic<uint32_t> *>(FirstAddress),
     82         uint32_t(PatchOpcodes::PO_StpX0X30SP_m16e), std::memory_order_release);
     83   } else {
     84     std::atomic_store_explicit(
     85         reinterpret_cast<std::atomic<uint32_t> *>(FirstAddress),
     86         uint32_t(PatchOpcodes::PO_B32), std::memory_order_release);
     87   }
     88   __clear_cache(reinterpret_cast<char *>(FirstAddress),
     89                 reinterpret_cast<char *>(CurAddress));
     90   return true;
     91 }
     92 
     93 bool patchFunctionEntry(const bool Enable, const uint32_t FuncId,
     94                         const XRaySledEntry &Sled,
     95                         void (*Trampoline)()) XRAY_NEVER_INSTRUMENT {
     96   return patchSled(Enable, FuncId, Sled, Trampoline);
     97 }
     98 
     99 bool patchFunctionExit(const bool Enable, const uint32_t FuncId,
    100                        const XRaySledEntry &Sled) XRAY_NEVER_INSTRUMENT {
    101   return patchSled(Enable, FuncId, Sled, __xray_FunctionExit);
    102 }
    103 
    104 bool patchFunctionTailExit(const bool Enable, const uint32_t FuncId,
    105                            const XRaySledEntry &Sled) XRAY_NEVER_INSTRUMENT {
    106   return patchSled(Enable, FuncId, Sled, __xray_FunctionTailExit);
    107 }
    108 
    109 bool patchCustomEvent(const bool Enable, const uint32_t FuncId,
    110                       const XRaySledEntry &Sled)
    111     XRAY_NEVER_INSTRUMENT { // FIXME: Implement in aarch64?
    112   return false;
    113 }
    114 
    115 bool patchTypedEvent(const bool Enable, const uint32_t FuncId,
    116                      const XRaySledEntry &Sled) XRAY_NEVER_INSTRUMENT {
    117   // FIXME: Implement in aarch64?
    118   return false;
    119 }
    120 
    121 // FIXME: Maybe implement this better?
    122 bool probeRequiredCPUFeatures() XRAY_NEVER_INSTRUMENT { return true; }
    123 
    124 } // namespace __xray
    125 
    126 extern "C" void __xray_ArgLoggerEntry() XRAY_NEVER_INSTRUMENT {
    127   // FIXME: this will have to be implemented in the trampoline assembly file
    128 }
    129