[LoongArch64] Fix some bugs to make NativeAOT stable.

src/coreclr/nativeaot/Common/src/Internal/Runtime/TransitionBlock.cs

@@ -312,8 +312,6 @@ internal struct ReturnBlock
     {
         private IntPtr returnValue;
         private IntPtr returnValue2;
-        private IntPtr returnValue3;
-        private IntPtr returnValue4;
     }
 
     [StructLayout(LayoutKind.Sequential)]
@@ -353,8 +351,8 @@ internal struct ArchitectureConstants
 
         public const int NUM_ARGUMENT_REGISTERS = 8;
         public const int ARGUMENTREGISTERS_SIZE = NUM_ARGUMENT_REGISTERS * 8;
-        public const int ENREGISTERED_RETURNTYPE_MAXSIZE = 32;                  // bytes (four FP registers: d0,d1,d2 and d3)
-        public const int ENREGISTERED_RETURNTYPE_INTEGER_MAXSIZE = 16;          // bytes (two int registers: x0 and x1)
+        public const int ENREGISTERED_RETURNTYPE_MAXSIZE = 16;                  // bytes (two FP registers: f0, f1)
+        public const int ENREGISTERED_RETURNTYPE_INTEGER_MAXSIZE = 16;          // bytes (two int registers: a0, a1)
         public const int ENREGISTERED_RETURNTYPE_INTEGER_MAXSIZE_PRIMITIVE = 8;
         public const int ENREGISTERED_PARAMTYPE_MAXSIZE = 16;                   // bytes (max value type size that can be passed by value)
         public const int STACK_ELEM_SIZE = 8;

src/coreclr/nativeaot/Runtime/CMakeLists.txt

@@ -245,7 +245,7 @@ endif()
 
 add_definitions(-DFEATURE_BASICFREEZE)
 add_definitions(-DFEATURE_CONSERVATIVE_GC)
-if(CLR_CMAKE_TARGET_ARCH_AMD64 OR CLR_CMAKE_TARGET_ARCH_ARM64 OR CLR_CMAKE_TARGET_ARCH_RISCV64)
+if(CLR_CMAKE_TARGET_ARCH_AMD64 OR CLR_CMAKE_TARGET_ARCH_ARM64 OR CLR_CMAKE_TARGET_ARCH_RISCV64 OR CLR_CMAKE_TARGET_ARCH_LOONGARCH64)
   add_definitions(-DFEATURE_USE_SOFTWARE_WRITE_WATCH_FOR_GC_HEAP)
   add_definitions(-DFEATURE_MANUALLY_MANAGED_CARD_BUNDLES)
 endif()

src/coreclr/nativeaot/Runtime/MiscHelpers.cpp

@@ -345,7 +345,7 @@ FCIMPL1(uint8_t *, RhGetCodeTarget, uint8_t * pCodeOrg)
         pCode++;
     }
     // is this an indirect jump?
-    // pcalau12i $t7, imm20; ld.d $t7, $t7, imm12; jirl $r0, $t7, 0
+    // pcalau12i $rd, imm20; ld.d $rd, $rj, imm12; jirl $rd, $rj, 0
     if ((pCode[0] & 0xfe000000) == 0x1a000000 &&
         (pCode[1] & 0xffc00000) == 0x28c00000 &&
         (pCode[2] & 0xfc000000) == 0x4c000000)

src/coreclr/nativeaot/Runtime/PalRedhawkCommon.h

@@ -115,7 +115,6 @@ struct PAL_LIMITED_CONTEXT
     uint64_t      F[32 - 24]; // Only the F registers F24..F31 need to be preserved
                               // (F0-F23 are not preserved according to the ABI spec).
 
-
     uintptr_t GetIp() const { return IP; }
     uintptr_t GetSp() const { return SP; }
     uintptr_t GetFp() const { return FP; }

src/coreclr/nativeaot/Runtime/StackFrameIterator.cpp

@@ -495,7 +495,7 @@ void StackFrameIterator::InternalInit(Thread * pThreadToWalk, PTR_PAL_LIMITED_CO
     m_RegDisplay.pRA = (PTR_uintptr_t)PTR_TO_MEMBER_TADDR(PAL_LIMITED_CONTEXT, pCtx, RA);
 
     //
-    // preserved vfp regs
+    // preserved fp regs
     //
     for (int32_t i = 0; i < 16 - 8; i++)
     {
@@ -1282,6 +1282,7 @@ void StackFrameIterator::UnwindFuncletInvokeThunk()
     m_RegDisplay.pR29 = SP++;
     m_RegDisplay.pR30 = SP++;
     m_RegDisplay.pR31 = SP++;
+    SP++; // for alignment padding
 
 #elif defined(TARGET_RISCV64)
     PTR_uint64_t f = (PTR_uint64_t)(m_RegDisplay.SP);
@@ -1476,12 +1477,12 @@ struct UniversalTransitionStackFrame
     // Conservative GC reporting must be applied to everything between the base of the
     // ReturnBlock and the top of the StackPassedArgs.
 private:
-    uintptr_t m_pushedFP;                  // ChildSP+000     CallerSP-0F0 (0x08 bytes)    (fp)
-    uintptr_t m_pushedRA;                  // ChildSP+008     CallerSP-0E8 (0x08 bytes)    (ra)
-    Fp128   m_fpArgRegs[8];                // ChildSP+010     CallerSP-0E0 (0x80 bytes)    (fa0-fa7)
-    uintptr_t m_returnBlock[4];            // ChildSP+090     CallerSP-060 (0x20 bytes)
-    uintptr_t m_intArgRegs[8];             // ChildSP+0B0     CallerSP-040 (0x40 bytes)    (a0-a7)
-    uintptr_t m_stackPassedArgs[1];        // ChildSP+0F0     CallerSP+000 (unknown size)
+    uintptr_t m_pushedFP;                  // ChildSP+000     CallerSP-0A0 (0x08 bytes)    (fp)
+    uintptr_t m_pushedRA;                  // ChildSP+008     CallerSP-098 (0x08 bytes)    (ra)
+    uint64_t  m_fpArgRegs[8];              // ChildSP+010     CallerSP-090 (0x40 bytes)    (fa0-fa7)
+    uintptr_t m_returnBlock[2];            // ChildSP+050     CallerSP-050 (0x10 bytes)
+    uintptr_t m_intArgRegs[8];             // ChildSP+060     CallerSP-040 (0x40 bytes)    (a0-a7)
+    uintptr_t m_stackPassedArgs[1];        // ChildSP+0A0     CallerSP+000 (unknown size)
 
 public:
     PTR_uintptr_t get_CallerSP() { return GET_POINTER_TO_FIELD(m_stackPassedArgs[0]); }

src/coreclr/nativeaot/Runtime/ThunksMapping.cpp

@@ -244,10 +244,14 @@ EXTERN_C void* QCALLTYPE RhAllocateThunksMapping()
             //jirl      $r0, $t8, 0
 
             int delta = (int)(pCurrentDataAddress - pCurrentThunkAddress);
+            ASSERT((-0x200000 <= delta) && (delta < 0x200000));
+
             *((uint32_t*)pCurrentThunkAddress) = 0x18000013 | (((delta & 0x3FFFFC) >> 2) << 5);
             pCurrentThunkAddress += 4;
 
             delta += OS_PAGE_SIZE - POINTER_SIZE - (i * POINTER_SIZE * 2) - 4;
+            ASSERT((-0x200000 <= delta) && (delta < 0x200000));
+
             *((uint32_t*)pCurrentThunkAddress) = 0x18000014 | (((delta & 0x3FFFFC) >> 2) << 5);
             pCurrentThunkAddress += 4;
 

src/coreclr/nativeaot/Runtime/gcenv.ee.cpp

@@ -399,7 +399,7 @@ void GCToEEInterface::StompWriteBarrier(WriteBarrierParameters* args)
         //     On architectures with strong ordering, we only need to prevent compiler reordering.
         //     Otherwise we put a process-wide fence here (so that we could use an ordinary read in the barrier)
 
-#if defined(HOST_ARM64) || defined(HOST_ARM)
+#if defined(HOST_ARM64) || defined(HOST_ARM) || defined(HOST_LOONGARCH64) || defined(HOST_RISCV64)
         if (!is_runtime_suspended)
         {
             // If runtime is not suspended, force all threads to see the changed table before seeing updated heap boundaries.
@@ -411,7 +411,7 @@ void GCToEEInterface::StompWriteBarrier(WriteBarrierParameters* args)
         g_lowest_address = args->lowest_address;
         g_highest_address = args->highest_address;
 
-#if defined(HOST_ARM64) || defined(HOST_ARM)
+#if defined(HOST_ARM64) || defined(HOST_ARM) || defined(HOST_LOONGARCH64) || defined(HOST_RISCV64)
         if (!is_runtime_suspended)
         {
             // If runtime is not suspended, force all threads to see the changed state before observing future allocations.

src/coreclr/nativeaot/Runtime/inc/rhbinder.h

@@ -341,8 +341,6 @@ enum PInvokeTransitionFrameFlags : uint64_t
 #elif defined(TARGET_LOONGARCH64)
 enum PInvokeTransitionFrameFlags : uint64_t
 {
-    // NOTE: Keep in sync with src\coreclr\nativeaot\Runtime\loongarch64\AsmMacros.h
-
     // NOTE: The order in which registers get pushed in the PInvokeTransitionFrame's m_PreservedRegs list has
     //       to match the order of these flags (that's also the order in which they are read in StackFrameIterator.cpp
 

src/coreclr/nativeaot/Runtime/loongarch64/AllocFast.S

@@ -13,8 +13,6 @@
 #define OFFSETOF__Thread__m_alloc_context__alloc_ptr        (OFFSETOF__Thread__m_eeAllocContext + OFFSETOF__ee_alloc_context__m_rgbAllocContextBuffer + OFFSETOF__gc_alloc_context__alloc_ptr)
 #define OFFSETOF__Thread__m_eeAllocContext__combined_limit  (OFFSETOF__Thread__m_eeAllocContext + OFFSETOF__ee_alloc_context__combined_limit)
 
-
-
 // Allocate non-array, non-finalizable object. If the allocation doesn't fit into the current thread's
 // allocation context then automatically fallback to the slow allocation path.
 //  $a0 == MethodTable
@@ -83,7 +81,7 @@ LOCAL_LABEL(RhpNewFast_RarePath):
         bl  C_FUNC(RhpGcAlloc)
 
         // Set the new objects MethodTable pointer on success.
-        beq  $a0, $zero, LOCAL_LABEL(NewOutOfMemory)
+        beqz  $a0, LOCAL_LABEL(NewOutOfMemory)
 
         .cfi_remember_state
         POP_COOP_PINVOKE_FRAME
@@ -245,7 +243,7 @@ LOCAL_LABEL(RhpNewArray_Rare):
         bl  C_FUNC(RhpGcAlloc)
 
         // Set the new objects MethodTable pointer and length on success.
-        beq  $a0, $zero, LOCAL_LABEL(ArrayOutOfMemory)
+        beqz  $a0, LOCAL_LABEL(ArrayOutOfMemory)
 
         .cfi_remember_state
         POP_COOP_PINVOKE_FRAME

src/coreclr/nativeaot/Runtime/loongarch64/AsmOffsetsCpu.h

@@ -13,7 +13,7 @@ PLAT_ASM_OFFSET(8, ExInfo, m_pExContext)
 PLAT_ASM_OFFSET(10, ExInfo, m_exception)
 PLAT_ASM_OFFSET(18, ExInfo, m_kind)
 PLAT_ASM_OFFSET(19, ExInfo, m_passNumber)
-PLAT_ASM_OFFSET(1c, ExInfo, m_idxCurClause)
+PLAT_ASM_OFFSET(1C, ExInfo, m_idxCurClause)
 PLAT_ASM_OFFSET(20, ExInfo, m_frameIter)
 PLAT_ASM_OFFSET(268, ExInfo, m_notifyDebuggerSP)
 
@@ -51,16 +51,16 @@ PLAT_ASM_OFFSET(70, PAL_LIMITED_CONTEXT, IP)
 PLAT_ASM_SIZEOF(148, REGDISPLAY)
 PLAT_ASM_OFFSET(18, REGDISPLAY, SP)
 
-PLAT_ASM_OFFSET(b8, REGDISPLAY, pR23)
-PLAT_ASM_OFFSET(c0, REGDISPLAY, pR24)
-PLAT_ASM_OFFSET(c8, REGDISPLAY, pR25)
-PLAT_ASM_OFFSET(d0, REGDISPLAY, pR26)
-PLAT_ASM_OFFSET(d8, REGDISPLAY, pR27)
-PLAT_ASM_OFFSET(e0, REGDISPLAY, pR28)
-PLAT_ASM_OFFSET(e8, REGDISPLAY, pR29)
-PLAT_ASM_OFFSET(f0, REGDISPLAY, pR30)
-PLAT_ASM_OFFSET(f8, REGDISPLAY, pR31)
+PLAT_ASM_OFFSET(B8, REGDISPLAY, pR23)
+PLAT_ASM_OFFSET(C0, REGDISPLAY, pR24)
+PLAT_ASM_OFFSET(C8, REGDISPLAY, pR25)
+PLAT_ASM_OFFSET(D0, REGDISPLAY, pR26)
+PLAT_ASM_OFFSET(D8, REGDISPLAY, pR27)
+PLAT_ASM_OFFSET(E0, REGDISPLAY, pR28)
+PLAT_ASM_OFFSET(E8, REGDISPLAY, pR29)
+PLAT_ASM_OFFSET(F0, REGDISPLAY, pR30)
+PLAT_ASM_OFFSET(F8, REGDISPLAY, pR31)
 PLAT_ASM_OFFSET(10, REGDISPLAY, pR2)
-PLAT_ASM_OFFSET(b0, REGDISPLAY, pFP)
+PLAT_ASM_OFFSET(B0, REGDISPLAY, pFP)
 PLAT_ASM_OFFSET(8, REGDISPLAY, pRA)
 PLAT_ASM_OFFSET(108, REGDISPLAY, F)

src/coreclr/nativeaot/Runtime/loongarch64/ExceptionHandling.S

@@ -18,14 +18,15 @@
         ori  $a3, $sp, 0
 
         // Setup a PAL_LIMITED_CONTEXT on the stack {
+        // Total stack: 0xC0 (0x68+0x58)
         .if \exceptionType == HARDWARE_EXCEPTION
-            addi.d  $sp, $sp, -0x50
-            .cfi_adjust_cfa_offset  0x50
+            addi.d  $sp, $sp, -0x58
+            .cfi_adjust_cfa_offset  0x58
             st.d  $a3, $sp, 0      // a3 is the SP and a1 is the IP of the fault site
             st.d  $a1, $sp, 8
         .else
-            PROLOG_STACK_ALLOC  0x50
-            .cfi_adjust_cfa_offset  0x50
+            PROLOG_STACK_ALLOC  0x58
+            .cfi_adjust_cfa_offset  0x58
             st.d  $a3, $sp, 0      // a3 is the SP and ra is the IP of the fault site
             st.d  $ra, $sp, 8
         .endif
@@ -37,6 +38,8 @@
         fst.d  $f29, $sp, 0x38
         fst.d  $f30, $sp, 0x40
         fst.d  $f31, $sp, 0x48
+        // Slot at $sp+0x50 is alignment padding
+
         PROLOG_SAVE_REG_PAIR_INDEXED  22, 1, 0x68
         st.d  $zero, $sp, 0x10 // locations reserved for return value, not used for exception handling
         st.d  $zero, $sp, 0x18
@@ -61,7 +64,7 @@
         // sp in fp. If sp is saved in fp in prolog then it is not expected that fp can change in the body
         // of method. However, this method needs to be able to change fp before calling funclet.
         // This is required to access locals in funclet.
-        PROLOG_SAVE_REG_PAIR_NO_FP_INDEXED  22, 1, 0x58
+        PROLOG_SAVE_REG_PAIR_NO_FP_INDEXED  22, 1, 0x60
         PROLOG_SAVE_REG_PAIR  23, 24, 0x10
         PROLOG_SAVE_REG_PAIR  25, 26, 0x20
         PROLOG_SAVE_REG_PAIR  27, 28, 0x30
@@ -92,7 +95,7 @@
         EPILOG_RESTORE_REG_PAIR  27, 28, 0x30
         EPILOG_RESTORE_REG_PAIR  29, 30, 0x40
         EPILOG_RESTORE_REG       31,     0x50
-        EPILOG_RESTORE_REG_PAIR_INDEXED  22, 1, 0x58
+        EPILOG_RESTORE_REG_PAIR_INDEXED  22, 1, 0x60
    .endm
 
 
@@ -160,7 +163,7 @@
         ld.d  $t3, \regdisplayReg, OFFSETOF__REGDISPLAY__pFP
         st.d  $fp, $t3, 0
         //
-        // store vfp preserved regs
+        // store fp preserved regs
         //
         addi.d  $t3, \regdisplayReg, OFFSETOF__REGDISPLAY__F
         fst.d  $f24, $t3, 0x00
@@ -288,7 +291,7 @@
         // where the tail-calling thread had saved RA, which may not match where we have saved RA.
 
         ld.d  $a1, $a2, OFFSETOF__Thread__m_pvHijackedReturnAddress
-        beq  $a1, $zero, LOCAL_LABEL(NotHijacked)
+        beqz  $a1, LOCAL_LABEL(NotHijacked)
 
         ld.d  $a3, $a2, OFFSETOF__Thread__m_ppvHijackedReturnAddressLocation
 
@@ -504,7 +507,7 @@ LOCAL_LABEL(NotHijacked):
 
 LOCAL_LABEL(PopExInfoLoop):
         ld.d  $a3, $a3, OFFSETOF__ExInfo__m_pPrevExInfo         // a3 <- next ExInfo
-        beq  $a3, $zero, LOCAL_LABEL(DonePopping)               // if (pExInfo == null) { we're done }
+        beqz  $a3, LOCAL_LABEL(DonePopping)                     // if (pExInfo == null) { we're done }
         blt  $a3, $a2, LOCAL_LABEL(PopExInfoLoop)               // if (pExInfo < resume SP} { keep going }
 
 LOCAL_LABEL(DonePopping):
@@ -513,10 +516,10 @@ LOCAL_LABEL(DonePopping):
         PREPARE_EXTERNAL_VAR_INDIRECT_W RhpTrapThreads, $a3
 
         bstrpick.d  $t7, $a3, TrapThreadsFlags_AbortInProgress_Bit, TrapThreadsFlags_AbortInProgress_Bit
-        beq  $t7, $zero, LOCAL_LABEL(NoAbort)
+        beqz  $t7, LOCAL_LABEL(NoAbort)
 
         ld.d  $a3, $sp, rsp_offset_is_not_handling_thread_abort
-        bne  $a3, $zero, LOCAL_LABEL(NoAbort)
+        bnez  $a3, LOCAL_LABEL(NoAbort)
 
         // It was the ThreadAbortException, so rethrow it
         // reset SP
@@ -773,7 +776,7 @@ LOCAL_LABEL(NoAbort):
 
 LOCAL_LABEL(Propagate_PopExInfoLoop):
         ld.d  $a3, $a3, OFFSETOF__ExInfo__m_pPrevExInfo     // a3 <- next ExInfo
-        beq  $a3, $zero, LOCAL_LABEL(Propagate_DonePopping) // if (pExInfo == null) { we're done }
+        beqz  $a3, LOCAL_LABEL(Propagate_DonePopping)       // if (pExInfo == null) { we're done }
         blt  $a3, $a2, LOCAL_LABEL(Propagate_PopExInfoLoop) // if (pExInfo < resume SP} { keep going }
 
 LOCAL_LABEL(Propagate_DonePopping):

src/coreclr/nativeaot/Runtime/loongarch64/GcProbe.S

@@ -4,12 +4,11 @@
 #include <unixasmmacros.inc>
 #include "AsmOffsets.inc"
 
-#define PROBE_FRAME_SIZE 0xC8  // 4 * 8  for fixed part of PInvokeTransitionFrame (fp, ra, m_pThread, m_Flags) +
-                               // 9 * 8 for callee saved registers +
+#define PROBE_FRAME_SIZE 0x90  // 4 * 8  for fixed part of PInvokeTransitionFrame (fp, ra, m_pThread, m_Flags) +
+                               // 9 * 8  for callee saved registers +
                                // 1 * 8  for caller SP +
                                // 2 * 8  for int returns +
-                               // 1 * 8  for alignment padding +
-                               // 4 * 16 for FP returns
+                               // 2 * 8  for FP returns
 
 // See PUSH_COOP_PINVOKE_FRAME, this macro is very similar, but also saves return registers
 // and accepts the register bitmask
@@ -42,13 +41,9 @@
     st.d  $a0, $sp, 0x70
     st.d  $a1, $sp, 0x78
 
-    // Slot at $sp+0x80 is alignment padding
-
     // Save the FP return registers
-    fst.d  $f0, $sp, 0x88
-    fst.d  $f1, $sp, 0x90
-    fst.d  $f2, $sp, 0x98
-    fst.d  $f3, $sp, 0xA0
+    fst.d  $f0, $sp, 0x80
+    fst.d  $f1, $sp, 0x88
 
     // Perform the rest of the PInvokeTransitionFrame initialization.
     st.d  \threadReg, $sp, OFFSETOF__PInvokeTransitionFrame__m_pThread   // Thread * (unused by stackwalker)
@@ -58,8 +53,7 @@
     st.d  \trashReg, $sp, 0x68                                           // save caller's SP
 
     // link the frame into the Thread
-    ori  \trashReg, $sp, 0
-    st.d  \trashReg, \threadReg, OFFSETOF__Thread__m_pDeferredTransitionFrame
+    st.d  $sp, \threadReg, OFFSETOF__Thread__m_pDeferredTransitionFrame
 .endm
 
 //
@@ -74,10 +68,8 @@
     ld.d  $a1, $sp, 0x78
 
     // Restore the FP return registers
-    fld.d  $f0, $sp, 0x88
-    fld.d  $f1, $sp, 0x90
-    fld.d  $f2, $sp, 0x98
-    fld.d  $f3, $sp, 0xA0
+    fld.d  $f0, $sp, 0x80
+    fld.d  $f1, $sp, 0x88
 
     // Restore callee saved registers
     EPILOG_RESTORE_REG_PAIR 23, 24, 0x20
@@ -126,12 +118,11 @@ NESTED_ENTRY RhpGcProbeHijack, _TEXT, NoHandler
 
     PREPARE_EXTERNAL_VAR_INDIRECT_W RhpTrapThreads, $a3
     bstrpick.d  $t8, $a3, TrapThreadsFlags_TrapThreads_Bit, TrapThreadsFlags_TrapThreads_Bit
-    bne  $t8, $zero, LOCAL_LABEL(WaitForGC)
+    bnez  $t8, LOCAL_LABEL(WaitForGC)
     jirl  $r0, $ra, 0
 
 LOCAL_LABEL(WaitForGC):
-    lu12i.w  $t3, ((DEFAULT_FRAME_SAVE_FLAGS + PTFF_SAVE_R4 + PTFF_SAVE_R5 + PTFF_THREAD_HIJACK_HI) >> 12) & 0xfffff
-    ori  $t3, $t3, (DEFAULT_FRAME_SAVE_FLAGS + PTFF_SAVE_R4 + PTFF_SAVE_R5 + PTFF_THREAD_HIJACK_HI) & 0xfff
+    li.d  $t3, (DEFAULT_FRAME_SAVE_FLAGS + PTFF_SAVE_R4 + PTFF_SAVE_R5 + (PTFF_THREAD_HIJACK_HI << 32))
     b  C_FUNC(RhpWaitForGC)
 NESTED_END RhpGcProbeHijack
 
@@ -145,7 +136,7 @@ NESTED_ENTRY RhpWaitForGC, _TEXT, NoHandler
 
     ld.d  $a2,$sp, OFFSETOF__PInvokeTransitionFrame__m_Flags
     bstrpick.d  $t8, $a2, PTFF_THREAD_ABORT_BIT, PTFF_THREAD_ABORT_BIT
-    bne  $t8, $zero, LOCAL_LABEL(ThrowThreadAbort)
+    bnez  $t8, LOCAL_LABEL(ThrowThreadAbort)
 
     .cfi_remember_state
     POP_PROBE_FRAME
@@ -163,7 +154,7 @@ NESTED_END RhpWaitForGC
 
 LEAF_ENTRY RhpGcPoll
     PREPARE_EXTERNAL_VAR_INDIRECT_W RhpTrapThreads, $a0
-    bne  $a0, $zero, C_FUNC(RhpGcPollRare)
+    bnez  $a0, C_FUNC(RhpGcPollRare)
     jirl  $r0, $ra, 0
 LEAF_END RhpGcPoll
 

src/coreclr/nativeaot/Runtime/loongarch64/InteropThunksHelpers.S

@@ -12,19 +12,19 @@
     //
     // RhCommonStub
     //
-    //  INPUT: tp: thunk's data block
+    //  INPUT: t7: thunk's data block
     //
-    //  TRASHES: t0, t1, tp
+    //  TRASHES: t0, t1, t7
     //
     LEAF_ENTRY RhCommonStub, _TEXT
         // There are arbitrary callers passing arguments with arbitrary signatures.
         // Custom calling convention:
-        //      tp pointer to the current thunk's data block (data contains 2 pointer values: context + target pointers)
+        //      t7 pointer to the current thunk's data block (data contains 2 pointer values: context + target pointers)
 
         INLINE_GET_TLS_VAR  $t0, C_FUNC(tls_thunkData)
 
         // t0  = base address of TLS data
-        // tp = address of context cell in thunk's data
+        // t7 = address of context cell in thunk's data
 
         // store thunk address in thread static
         ld.d  $t1, $t7, 0