/* * Copyright (C) 2023 The Android Open Source Project * * Licensed under the Apache License, Version 2.0 (the "License"); * you may not use this file except in compliance with the License. * You may obtain a copy of the License at * * http://www.apache.org/licenses/LICENSE-2.0 * * Unless required by applicable law or agreed to in writing, software * distributed under the License is distributed on an "AS IS" BASIS, * WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. * See the License for the specific language governing permissions and * limitations under the License. */ #include "asm_support_riscv64.S" .macro JNI_SAVE_MANAGED_ARGS_TRAMPOLINE name, cxx_name, arg1 = "none" .extern \cxx_name ENTRY \name // Save args and RA. SAVE_ALL_ARGS_INCREASE_FRAME /*padding*/ 8 + /*RA*/ 8 SAVE_GPR ra, (ALL_ARGS_SIZE + /*padding*/ 8) // Call `cxx_name()`. .ifnc \arg1, none mv a0, \arg1 .endif call \cxx_name // Restore RA and args and return. RESTORE_GPR ra, (ALL_ARGS_SIZE + /*padding*/ 8) RESTORE_ALL_ARGS_DECREASE_FRAME /*padding*/ 8 + /*RA*/ 8 ret END \name .endm .macro JNI_SAVE_RETURN_VALUE_TRAMPOLINE name, cxx_name, arg1, arg2 = "none" .extern \cxx_name ENTRY \name // Save return registers and return address. INCREASE_FRAME 32 sd a0, 0(sp) fsd fa0, 8(sp) SAVE_GPR ra, 24 // Call `cxx_name()`. mv a0, \arg1 .ifnc \arg2, none mv a1, \arg2 .endif call \cxx_name // Restore result registers and return. ld a0, 0(sp) fld fa0, 8(sp) RESTORE_GPR ra, 24 DECREASE_FRAME 32 ret END \name .endm // JNI dlsym lookup stub for @CriticalNative. ENTRY art_jni_dlsym_lookup_critical_stub // The hidden arg holding the tagged method is t0 (loaded by compiled JNI stub, compiled // managed code, or `art_quick_generic_jni_trampoline`). Bit 0 set means generic JNI. // For generic JNI we already have a managed frame, so we reuse the art_jni_dlsym_lookup_stub. // Note: 'bnez' doesn't always have enough range (+/-4KB) to reach art_jni_dlsym_lookup_stub so // 'tail' is used instead. andi t6, t0, 1 beqz t6, .Lcritical_not_generic_jni tail art_jni_dlsym_lookup_stub .Lcritical_not_generic_jni: // Save args, the hidden arg and caller PC. No CFI needed for args and the hidden arg. SAVE_ALL_ARGS_INCREASE_FRAME 2*8 SAVE_GPR t0, (ALL_ARGS_SIZE + 0) SAVE_GPR ra, (ALL_ARGS_SIZE + 8) // Call artCriticalNativeFrameSize(method, caller_pc) mv a0, t0 // a0 := method (from hidden arg) mv a1, ra // a1 := caller_pc call artCriticalNativeFrameSize // Move frame size to T2. mv t2, a0 // Restore args, the hidden arg and caller PC. RESTORE_GPR t0, (ALL_ARGS_SIZE + 0) RESTORE_GPR ra, (ALL_ARGS_SIZE + 8) RESTORE_ALL_ARGS_DECREASE_FRAME 2*8 // Reserve space for a SaveRefsAndArgs managed frame, either for the actual runtime // method or for a GenericJNI frame which is similar but has a native method and a tag. // Add space for RA and padding to keep the stack 16-byte aligned. INCREASE_FRAME (FRAME_SIZE_SAVE_REFS_AND_ARGS + 16) // Prepare the return address for managed stack walk of the SaveRefsAndArgs frame. // If we're coming from JNI stub with tail call, it is RA. If we're coming from // JNI stub that saved the return address, it will be the last value we copy below. // If we're coming directly from compiled code, it is RA, set further down. mv t4, ra // Move the stack args if any. Calculate the base address of the managed frame in the process. addi t1, sp, 16 beqz t2, .Lcritical_skip_copy_args .Lcritical_copy_args_loop: ld t3, FRAME_SIZE_SAVE_REFS_AND_ARGS+0(t1) ld t4, FRAME_SIZE_SAVE_REFS_AND_ARGS+8(t1) addi t2, t2, -16 sd t3, 0-16(t1) sd t4, 8-16(t1) addi t1, t1, 16 bnez t2, .Lcritical_copy_args_loop .Lcritical_skip_copy_args: // Spill registers for the SaveRefsAndArgs frame above the stack args. // Note that the runtime shall not examine the args here, otherwise we would have to // move them in registers and stack to account for the difference between managed and // native ABIs. Do not update CFI while we hold the frame address in T1 and the values // in registers are unchanged. // stack slot (0*8)(t1) is for ArtMethod* fsd fa0, (1*8)(t1) fsd fa1, (2*8)(t1) fsd fa2, (3*8)(t1) fsd fa3, (4*8)(t1) fsd fa4, (5*8)(t1) fsd fa5, (6*8)(t1) fsd fa6, (7*8)(t1) fsd fa7, (8*8)(t1) sd fp, (9*8)(t1) // x8, frame pointer // s1 (x9) is the ART thread register // a0 (x10) is the method pointer sd a1, (10*8)(t1) // x11 sd a2, (11*8)(t1) // x12 sd a3, (12*8)(t1) // x13 sd a4, (13*8)(t1) // x14 sd a5, (14*8)(t1) // x15 sd a6, (15*8)(t1) // x16 sd a7, (16*8)(t1) // x17 sd s2, (17*8)(t1) // x18 sd s3, (18*8)(t1) // x19 sd s4, (19*8)(t1) // x20 sd s5, (20*8)(t1) // x21 sd s6, (21*8)(t1) // x22 sd s7, (22*8)(t1) // x23 sd s8, (23*8)(t1) // x24 sd s9, (24*8)(t1) // x25 sd s10, (25*8)(t1) // x26 sd s11, (26*8)(t1) // x27 sd t4, (27*8)(t1) // t4: Save return address for tail call from JNI stub. // (If there were any stack args, we're storing the value that's already there. // For direct calls from compiled managed code, we shall overwrite this below.) // Move the managed frame address to native callee-save register fp (x8) and update CFI. mv fp, t1 // Skip args FA0-FA7, A1-A7 CFI_EXPRESSION_BREG 8, 8, (9*8) CFI_EXPRESSION_BREG 18, 8, (17*8) CFI_EXPRESSION_BREG 19, 8, (18*8) CFI_EXPRESSION_BREG 20, 8, (19*8) CFI_EXPRESSION_BREG 21, 8, (20*8) CFI_EXPRESSION_BREG 22, 8, (21*8) CFI_EXPRESSION_BREG 23, 8, (22*8) CFI_EXPRESSION_BREG 24, 8, (23*8) CFI_EXPRESSION_BREG 25, 8, (24*8) CFI_EXPRESSION_BREG 26, 8, (25*8) CFI_EXPRESSION_BREG 27, 8, (26*8) // The saved return PC for managed stack walk is not necessarily our RA. // Save our return PC below the managed frame. sd ra, -__SIZEOF_POINTER__(fp) CFI_EXPRESSION_BREG 1, 8, -__SIZEOF_POINTER__ lw t2, ART_METHOD_ACCESS_FLAGS_OFFSET(t0) // Load access flags. addi t1, fp, 1 // Prepare managed SP tagged for a GenericJNI frame. slliw t2, t2, 31 - ACCESS_FLAGS_METHOD_IS_NATIVE_BIT bltz t2, .Lcritical_skip_prepare_runtime_method // When coming from a compiled method, the return PC for managed stack walk is RA. // (When coming from a compiled stub, the correct return PC is already stored above.) sd ra, (FRAME_SIZE_SAVE_REFS_AND_ARGS - __SIZEOF_POINTER__)(fp) // Replace the target method with the SaveRefsAndArgs runtime method. LOAD_RUNTIME_INSTANCE t0 ld t0, RUNTIME_SAVE_REFS_AND_ARGS_METHOD_OFFSET(t0) mv t1, fp // Prepare untagged managed SP for the runtime method. .Lcritical_skip_prepare_runtime_method: // Store the method on the bottom of the managed frame. sd t0, (fp) // Place (maybe tagged) managed SP in Thread::Current()->top_quick_frame. sd t1, THREAD_TOP_QUICK_FRAME_OFFSET(xSELF) // Preserve the native arg register A0 in callee-save register S2 (x18) which was saved above. mv s2, a0 // Call artFindNativeMethodRunnable() mv a0, xSELF // pass Thread::Current() call artFindNativeMethodRunnable // Store result in scratch reg. mv t0, a0 // Restore the native arg register A0. mv a0, s2 // Restore our return PC. RESTORE_GPR_BASE fp, ra, -__SIZEOF_POINTER__ // Remember the end of out args before restoring FP. addi t1, fp, -16 // Restore arg registers. fld fa0, (1*8)(fp) fld fa1, (2*8)(fp) fld fa2, (3*8)(fp) fld fa3, (4*8)(fp) fld fa4, (5*8)(fp) fld fa5, (6*8)(fp) fld fa6, (7*8)(fp) fld fa7, (8*8)(fp) // fp (x8) is restored last to keep CFI data valid until then. // s1 (x9) is the ART thread register // a0 (x10) is the method pointer ld a1, (10*8)(fp) // x11 ld a2, (11*8)(fp) // x12 ld a3, (12*8)(fp) // x13 ld a4, (13*8)(fp) // x14 ld a5, (14*8)(fp) // x15 ld a6, (15*8)(fp) // x16 ld a7, (16*8)(fp) // x17 RESTORE_GPR_BASE fp, s2, (17*8) // x18 RESTORE_GPR_BASE fp, s3, (18*8) // x19 RESTORE_GPR_BASE fp, s4, (19*8) // x20 RESTORE_GPR_BASE fp, s5, (20*8) // x21 RESTORE_GPR_BASE fp, s6, (21*8) // x22 RESTORE_GPR_BASE fp, s7, (22*8) // x23 RESTORE_GPR_BASE fp, s8, (23*8) // x24 RESTORE_GPR_BASE fp, s9, (24*8) // x25 RESTORE_GPR_BASE fp, s10, (25*8) // x26 RESTORE_GPR_BASE fp, s11, (26*8) // x27 RESTORE_GPR_BASE fp, fp, (9*8) // fp (x8) is restored last // Check for exception before moving args back to keep the return PC for managed stack walk. CFI_REMEMBER_STATE beqz t0, .Lcritical_deliver_exception // Move stack args to their original place. beq t1, sp, .Lcritical_skip_copy_args_back sub t2, t1, sp .Lcritical_copy_args_back_loop: ld t3, 0-16(t1) ld t4, 8-16(t1) addi t2, t2, -16 sd t3, FRAME_SIZE_SAVE_REFS_AND_ARGS+0(t1) sd t4, FRAME_SIZE_SAVE_REFS_AND_ARGS+8(t1) addi t1, t1, -16 bnez t2, .Lcritical_copy_args_back_loop .Lcritical_skip_copy_args_back: // Remove the frame reservation. DECREASE_FRAME (FRAME_SIZE_SAVE_REFS_AND_ARGS + 16) // Do the tail call. jr t0 .Lcritical_deliver_exception: CFI_RESTORE_STATE_AND_DEF_CFA sp, FRAME_SIZE_SAVE_REFS_AND_ARGS + 16 // If this is called from a method that catches the exception, all callee-save registers need // to be saved, so that the exception handling code can read them in case they contain live // values later used by that method. This includes callee-save FP registers which are not // saved in a SaveRefsAndArgs frame, so we cannot reuse the managed frame we have built above. // That's why we checked for exception after restoring registers from that frame. // We need to build a SaveAllCalleeSaves frame instead. Args are irrelevant at this // point but keep the area allocated for stack args to keep CFA definition simple. #if FRAME_SIZE_SAVE_ALL_CALLEE_SAVES > FRAME_SIZE_SAVE_REFS_AND_ARGS #error "Expanding stack frame from kSaveRefsAndArgs to kSaveAllCalleeSaves is not implemented." #endif DECREASE_FRAME FRAME_SIZE_SAVE_REFS_AND_ARGS - FRAME_SIZE_SAVE_ALL_CALLEE_SAVES // Calculate the base address of the managed frame. addi t1, t1, 16 + FRAME_SIZE_SAVE_REFS_AND_ARGS - FRAME_SIZE_SAVE_ALL_CALLEE_SAVES // Spill registers for the SaveAllCalleeSaves frame above the stack args area. Do not update // CFI while we hold the frame address in T1 and the values in registers are unchanged. // stack slot (0*8)(t1) is for ArtMethod* // stack slot (1*8)(t1) is for padding // FP callee-saves. fsd fs0, (8*2)(t1) // f8 fsd fs1, (8*3)(t1) // f9 fsd fs2, (8*4)(t1) // f18 fsd fs3, (8*5)(t1) // f19 fsd fs4, (8*6)(t1) // f20 fsd fs5, (8*7)(t1) // f21 fsd fs6, (8*8)(t1) // f22 fsd fs7, (8*9)(t1) // f23 fsd fs8, (8*10)(t1) // f24 fsd fs9, (8*11)(t1) // f25 fsd fs10, (8*12)(t1) // f26 fsd fs11, (8*13)(t1) // f27 // GP callee-saves sd s0, (8*14)(t1) // x8/fp, frame pointer // s1 (x9) is the ART thread register sd s2, (8*15)(t1) // x18 sd s3, (8*16)(t1) // x19 sd s4, (8*17)(t1) // x20 sd s5, (8*18)(t1) // x21 sd s6, (8*19)(t1) // x22 sd s7, (8*20)(t1) // x23 sd s8, (8*21)(t1) // x24 sd s9, (8*22)(t1) // x25 sd s10, (8*23)(t1) // x26 sd s11, (8*24)(t1) // x27 // Keep the caller PC for managed stack walk. // Move the managed frame address to native callee-save register fp (x8) and update CFI. mv fp, t1 CFI_EXPRESSION_BREG 8, 8, (14*8) // fp/x8: The base register for these CFI expressions. CFI_EXPRESSION_BREG /*FP reg*/ 32 + 8, 8, (8*2) // fs0/f8 CFI_EXPRESSION_BREG /*FP reg*/ 32 + 9, 8, (8*3) // fs1/f9 CFI_EXPRESSION_BREG /*FP reg*/ 32 + 18, 8, (8*4) // fs2/f18 CFI_EXPRESSION_BREG /*FP reg*/ 32 + 19, 8, (8*5) // fs3/f19 CFI_EXPRESSION_BREG /*FP reg*/ 32 + 20, 8, (8*6) // fs4/f20 CFI_EXPRESSION_BREG /*FP reg*/ 32 + 21, 8, (8*7) // fs5/f21 CFI_EXPRESSION_BREG /*FP reg*/ 32 + 22, 8, (8*8) // fs6/f22 CFI_EXPRESSION_BREG /*FP reg*/ 32 + 23, 8, (8*9) // fs7/f23 CFI_EXPRESSION_BREG /*FP reg*/ 32 + 24, 8, (8*10) // fs8/f24 CFI_EXPRESSION_BREG /*FP reg*/ 32 + 25, 8, (8*11) // fs9/f25 CFI_EXPRESSION_BREG /*FP reg*/ 32 + 26, 8, (8*12) // fs10/f26 // CFI expression for fp (x8) already emitted above. CFI_EXPRESSION_BREG 18, 8, (15*8) // s2/x18 CFI_EXPRESSION_BREG 19, 8, (16*8) // s3/x19 CFI_EXPRESSION_BREG 20, 8, (17*8) // s4/x20 CFI_EXPRESSION_BREG 21, 8, (18*8) // s5/x21 CFI_EXPRESSION_BREG 22, 8, (19*8) // s6/x22 CFI_EXPRESSION_BREG 23, 8, (20*8) // s7/x23 CFI_EXPRESSION_BREG 24, 8, (21*8) // s8/x24 CFI_EXPRESSION_BREG 25, 8, (22*8) // s9/x25 CFI_EXPRESSION_BREG 26, 8, (23*8) // s10/x26 CFI_EXPRESSION_BREG 27, 8, (24*8) // s11/x27 // The saved return PC for managed stack walk is not necessarily our RA. // Save our return PC below the managed frame. sd ra, -__SIZEOF_POINTER__(fp) CFI_EXPRESSION_BREG 1, 8, -__SIZEOF_POINTER__ // Store ArtMethod* Runtime::callee_save_methods_[kSaveAllCalleeSaves] to the managed frame. LOAD_RUNTIME_INSTANCE t0 ld t0, RUNTIME_SAVE_ALL_CALLEE_SAVES_METHOD_OFFSET(t0) sd t0, (fp) // Place the managed frame SP in Thread::Current()->top_quick_frame. sd fp, THREAD_TOP_QUICK_FRAME_OFFSET(xSELF) DELIVER_PENDING_EXCEPTION_FRAME_READY END art_jni_dlsym_lookup_critical_stub /* * Read barrier for the method's declaring class needed by JNI stub for static methods. * (We're using a pointer to the declaring class in `ArtMethod` as `jclass`.) */ // The method argument is already in a0 for call to `artJniReadBarrier(ArtMethod*)`. JNI_SAVE_MANAGED_ARGS_TRAMPOLINE art_jni_read_barrier, artJniReadBarrier /* * Trampoline to `artJniMethodStart()` that preserves all managed arguments. */ JNI_SAVE_MANAGED_ARGS_TRAMPOLINE art_jni_method_start, artJniMethodStart, xSELF /* * Trampoline to `artJniMethodEntryHook` that preserves all managed arguments. */ JNI_SAVE_MANAGED_ARGS_TRAMPOLINE art_jni_method_entry_hook, artJniMethodEntryHook, xSELF /* * Trampoline to `artJniMonitoredMethodStart()` that preserves all managed arguments. */ JNI_SAVE_MANAGED_ARGS_TRAMPOLINE art_jni_monitored_method_start, artJniMonitoredMethodStart, xSELF /* * Trampoline to `artJniMethodEnd()` that preserves all return registers. */ JNI_SAVE_RETURN_VALUE_TRAMPOLINE art_jni_method_end, artJniMethodEnd, xSELF /* * Trampoline to `artJniMonitoredMethodEnd()` that preserves all return registers. */ JNI_SAVE_RETURN_VALUE_TRAMPOLINE art_jni_monitored_method_end, artJniMonitoredMethodEnd, xSELF /* * Entry from JNI stub that tries to lock the object in a fast path and * calls `artLockObjectFromCode()` (the same as for managed code) for the * difficult cases, may block for GC. * Custom calling convention: * T0 holds the non-null object to lock. * Callee-save registers have been saved and can be used as temporaries. * All argument registers need to be preserved. */ ENTRY art_jni_lock_object LOCK_OBJECT_FAST_PATH t0, art_jni_lock_object_no_inline, /*can_be_null*/ 0 END art_jni_lock_object /* * Entry from JNI stub that calls `artLockObjectFromCode()` * (the same as for managed code), may block for GC. * Custom calling convention: * T0 holds the non-null object to lock. * Callee-save registers have been saved and can be used as temporaries. * All argument registers need to be preserved. */ .extern artLockObjectFromCode ENTRY art_jni_lock_object_no_inline // This is also the slow path for art_jni_lock_object. // Save args and RA. SAVE_ALL_ARGS_INCREASE_FRAME /*padding*/ 8 + /*RA*/ 8 SAVE_GPR ra, (ALL_ARGS_SIZE + /*padding*/ 8) // Call `artLockObjectFromCode()`. mv a0, t0 // Pass the object to lock. mv a1, xSELF // Pass Thread::Current(). call artLockObjectFromCode // (Object* obj, Thread*) // Restore return address. RESTORE_GPR ra, (ALL_ARGS_SIZE + /*padding*/ 8) // Check result. bnez a0, 1f // Restore register args a0-a7, fa0-fa7 and return. RESTORE_ALL_ARGS_DECREASE_FRAME /*padding*/ 8 + /*RA*/ 8 ret .cfi_adjust_cfa_offset (ALL_ARGS_SIZE + /*padding*/ 8 + /*RA*/ 8) 1: // All args are irrelevant when throwing an exception. Remove the spill area. DECREASE_FRAME (ALL_ARGS_SIZE + /*padding*/ 8 + /*RA*/ 8) // Make a call to `artDeliverPendingExceptionFromCode()`. // Rely on the JNI transition frame constructed in the JNI stub. mv a0, xSELF // Pass Thread::Current(). call artDeliverPendingExceptionFromCode // (Thread*) call art_quick_do_long_jump // (Context*) unimp // Unreached END art_jni_lock_object_no_inline /* * Entry from JNI stub that tries to unlock the object in a fast path and calls * `artJniUnlockObject()` for the difficult cases. Note that failure to unlock * is fatal, so we do not need to check for exceptions in the slow path. * Custom calling convention: * T0 holds the non-null object to unlock. * Callee-save registers have been saved and can be used as temporaries. * Return registers a0 and fa0 need to be preserved. */ ENTRY art_jni_unlock_object UNLOCK_OBJECT_FAST_PATH t0, art_jni_unlock_object_no_inline, /*can_be_null*/ 0 END art_jni_unlock_object /* * Entry from JNI stub that calls `artJniUnlockObject()`. Note that failure to * unlock is fatal, so we do not need to check for exceptions. * Custom calling convention: * T0 holds the non-null object to unlock. * Callee-save registers have been saved and can be used as temporaries. * Return registers a0 and fa0 need to be preserved. */ // This is also the slow path for art_jni_unlock_object. JNI_SAVE_RETURN_VALUE_TRAMPOLINE art_jni_unlock_object_no_inline, artJniUnlockObject, t0, xSELF